Adding missed files

Adds updated_at TIMESTAMPTZ to ocr_jobs, stamped automatically by a
BEFORE UPDATE trigger. The sync process can use updated_at > qdrant_synced_at
to detect jobs that need re-ingestion after edits or reviews.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

.env

@@ -0,0 +1,23 @@
+# DeepSeek OCR Application Configuration
+
+# API Configuration
+API_HOST=0.0.0.0
+API_PORT=8000
+
+# Frontend Configuration
+FRONTEND_PORT=3000
+
+# Model Configuration
+MODEL_NAME=deepseek-ai/DeepSeek-OCR
+HF_HOME=/models
+
+# CORS Configuration (comma-separated origins, defaults to http://localhost:3000)
+CORS_ORIGINS=http://localhost:3000
+
+# Upload Configuration
+MAX_UPLOAD_SIZE_MB=100
+
+# Processing Configuration
+BASE_SIZE=1024
+IMAGE_SIZE=640
+CROP_MODE=true

.env.example

@@ -11,12 +11,34 @@ FRONTEND_PORT=3000
 MODEL_NAME=deepseek-ai/DeepSeek-OCR
 HF_HOME=/models
 
+# OCR model selection
+# Register the local DeepSeek-OCR model (set to false for an Ollama-only deployment)
+ENABLE_DEEPSEEK_LOCAL=true
+# External Ollama host the backend should call (no trailing slash)
+OLLAMA_BASE_URL=http://host.docker.internal:11434
+# Comma-separated Ollama vision model tags to surface in the UI.
+# Pull these on the Ollama host first, e.g. `ollama pull glm-ocr`.
+OLLAMA_MODELS=glm-ocr,llama3.2-vision,minicpm-v,qwen2.5vl
+# Default model id selected in the UI (deepseek-local or ollama:<tag>)
+DEFAULT_OCR_MODEL=deepseek-local
+# Per-request timeout (seconds) for Ollama calls
+OLLAMA_TIMEOUT=300
+
 # CORS Configuration (comma-separated origins, defaults to http://localhost:3000)
 CORS_ORIGINS=http://localhost:3000
 
 # Upload Configuration
 MAX_UPLOAD_SIZE_MB=100
 
+# PostgreSQL Configuration
+POSTGRES_USER=ocr_user
+POSTGRES_PASSWORD=ocr_password
+POSTGRES_DB=ocr_db
+DATABASE_URL=postgresql://ocr_user:ocr_password@postgres:5432/ocr_db
+
+# OCR Image Storage (host path mounted into container)
+OCR_IMAGES_DIR=/data/ocr_images
+
 # Processing Configuration
 BASE_SIZE=1024
 IMAGE_SIZE=640

.gitignore

@@ -46,7 +46,7 @@ yarn.lock
 pnpm-lock.yaml
 
 # Environment
-.env
+#.env
 .env.local
 .env.development.local
 .env.test.local

README.md

@@ -172,6 +172,13 @@ FRONTEND_PORT=3000
 MODEL_NAME=deepseek-ai/DeepSeek-OCR
 HF_HOME=/models
 
+# OCR model selection (DeepSeek + Ollama)
+ENABLE_DEEPSEEK_LOCAL=true                          # register the local GPU model
+OLLAMA_BASE_URL=http://host.docker.internal:11434   # external Ollama host
+OLLAMA_MODELS=glm-ocr,llama3.2-vision,minicpm-v,qwen2.5vl
+DEFAULT_OCR_MODEL=deepseek-local                    # deepseek-local or ollama:<tag>
+OLLAMA_TIMEOUT=300                                  # per-request timeout (seconds)
+
 # Upload Configuration
 MAX_UPLOAD_SIZE_MB=100  # Maximum file upload size
 
@@ -186,13 +193,47 @@ CROP_MODE=true         # Enable dynamic cropping for large images
 - `API_HOST`: Backend API host (default: 0.0.0.0)
 - `API_PORT`: Backend API port (default: 8000)
 - `FRONTEND_PORT`: Frontend port (default: 3000)
-- `MODEL_NAME`: HuggingFace model identifier
+- `MODEL_NAME`: HuggingFace model identifier for the local DeepSeek-OCR model
 - `HF_HOME`: Model cache directory
+- `ENABLE_DEEPSEEK_LOCAL`: Register the local DeepSeek-OCR model (set `false` for an Ollama-only deployment with no GPU model loaded)
+- `OLLAMA_BASE_URL`: URL of an external Ollama server the backend calls for non-DeepSeek models
+- `OLLAMA_MODELS`: Comma-separated Ollama vision model tags to expose in the UI (pull them on the Ollama host first, e.g. `ollama pull glm-ocr`)
+- `DEFAULT_OCR_MODEL`: Model id selected by default (`deepseek-local` or `ollama:<tag>`)
+- `OLLAMA_TIMEOUT`: Per-request timeout in seconds for Ollama calls
 - `MAX_UPLOAD_SIZE_MB`: Maximum file upload size in megabytes
 - `BASE_SIZE`: Base image processing size (affects memory usage)
 - `IMAGE_SIZE`: Tile size for dynamic cropping
 - `CROP_MODE`: Enable/disable dynamic image cropping
 
+### Choosing an OCR Model
+
+The **Model** selector (next to the Mode selector) chooses which backend runs the OCR:
+
+- **DeepSeek-OCR (local GPU)** — the default. Loaded lazily on first use. Supports
+  every mode including grounding/bounding-box modes (Find), plus the Advanced
+  Settings (base size, crop mode, etc.).
+- **Ollama models** — any vision model pulled on your Ollama host and listed in
+  `OLLAMA_MODELS` (e.g. `glm-ocr`, `llama3.2-vision`). These run remotely on the
+  Ollama server. They return **plain text only**: bounding boxes are not produced,
+  so grounding modes (Find) and the DeepSeek-specific Advanced Settings are ignored
+  / disabled when an Ollama model is selected.
+
+Setup for Ollama models:
+
+```bash
+# On the machine running Ollama
+ollama pull glm-ocr
+ollama pull llama3.2-vision
+
+# Point the backend at it (in .env), then restart
+OLLAMA_BASE_URL=http://host.docker.internal:11434
+OLLAMA_MODELS=glm-ocr,llama3.2-vision
+```
+
+`GET /api/models` returns the registered models and their capabilities; the UI
+populates the selector from it. The model used for each job is stored on the job
+record (`ocr_model`) and shown in the Browse Jobs view.
+
 ## Tech Stack
 
 ### Frontend
@@ -377,6 +418,7 @@ For large images, the model uses dynamic cropping:
 
 **Parameters:**
 - `image` (file, required) - Image file to process (up to 100MB)
+- `model` (string) - OCR model id from `GET /api/models` (default: registry default). Grounding/Advanced settings apply to DeepSeek only.
 - `mode` (string) - OCR mode: `plain_ocr` | `describe` | `find_ref` | `freeform`
 - `prompt` (string) - Custom prompt for freeform mode
 - `grounding` (bool) - Enable bounding boxes (auto-enabled for find_ref)
@@ -416,6 +458,7 @@ Process PDF documents with OCR and export to various formats.
 
 **Parameters:**
 - `pdf_file` (file, required) - PDF file to process (up to 100MB)
+- `model` (string) - OCR model id from `GET /api/models` (default: registry default)
 - `mode` (string) - OCR mode: `plain_ocr` | `describe` | `find_ref` | `freeform`
 - `prompt` (string) - Custom prompt for freeform mode
 - `output_format` (string) - Output format: `markdown` | `html` | `docx` | `json`

backend/database.py

@@ -0,0 +1,115 @@
+import os
+import psycopg2
+import psycopg2.extras
+from contextlib import contextmanager
+from decouple import config as env_config
+
+DATABASE_URL = env_config(
+    "DATABASE_URL",
+    default="postgresql://ocr_user:ocr_password@postgres:5432/ocr_db"
+)
+
+
+def _get_conn():
+    return psycopg2.connect(DATABASE_URL, cursor_factory=psycopg2.extras.RealDictCursor)
+
+
+def init_db():
+    """Create tables if they don't exist. Called once at startup."""
+    conn = None
+    try:
+        conn = _get_conn()
+        with conn.cursor() as cur:
+            cur.execute("""
+                CREATE TABLE IF NOT EXISTS ocr_jobs (
+                    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+                    author TEXT,
+                    book TEXT,
+                    chapter TEXT,
+                    page TEXT,
+                    submitted_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
+                    image_path TEXT NOT NULL,
+                    original_filename TEXT,
+                    ocr_text TEXT,
+                    status TEXT NOT NULL DEFAULT 'unreviewed',
+                    reviewed_text TEXT,
+                    reviewer_name TEXT,
+                    reviewed_at TIMESTAMPTZ,
+                    mode TEXT
+                )
+            """)
+            # Index for fast full-text-style searches on common fields
+            cur.execute("""
+                CREATE INDEX IF NOT EXISTS ocr_jobs_status_idx ON ocr_jobs(status)
+            """)
+            cur.execute("""
+                CREATE INDEX IF NOT EXISTS ocr_jobs_submitted_at_idx ON ocr_jobs(submitted_at DESC)
+            """)
+            # Add columns introduced after initial schema (safe to run repeatedly)
+            cur.execute("""
+                ALTER TABLE ocr_jobs
+                ADD COLUMN IF NOT EXISTS describe_text TEXT
+            """)
+            cur.execute("""
+                ALTER TABLE ocr_jobs
+                ADD COLUMN IF NOT EXISTS freeform_text TEXT
+            """)
+            cur.execute("""
+                ALTER TABLE ocr_jobs
+                ADD COLUMN IF NOT EXISTS qdrant_synced_at TIMESTAMPTZ
+            """)
+            cur.execute("""
+                ALTER TABLE ocr_jobs
+                ADD COLUMN IF NOT EXISTS updated_at TIMESTAMPTZ
+            """)
+            # Which OCR model produced this job (e.g. "deepseek-local", "ollama:glm-ocr")
+            cur.execute("""
+                ALTER TABLE ocr_jobs
+                ADD COLUMN IF NOT EXISTS ocr_model TEXT
+            """)
+            # Trigger function: stamp updated_at on every row update
+            cur.execute("""
+                CREATE OR REPLACE FUNCTION set_updated_at()
+                RETURNS TRIGGER AS $$
+                BEGIN
+                    NEW.updated_at = NOW();
+                    RETURN NEW;
+                END;
+                $$ LANGUAGE plpgsql
+            """)
+            cur.execute("""
+                CREATE OR REPLACE TRIGGER ocr_jobs_set_updated_at
+                BEFORE UPDATE ON ocr_jobs
+                FOR EACH ROW EXECUTE FUNCTION set_updated_at()
+            """)
+            # Unique constraint: prevent duplicate (author, chapter, page) submissions.
+            # Applies only when all three fields are non-null.
+            cur.execute("""
+                CREATE UNIQUE INDEX IF NOT EXISTS ocr_jobs_author_chapter_page_unique
+                ON ocr_jobs (author, chapter, page)
+                WHERE author IS NOT NULL AND chapter IS NOT NULL AND page IS NOT NULL
+            """)
+        conn.commit()
+        print("Database initialized.")
+    except Exception as exc:
+        print(f"Database init failed: {exc}")
+        if conn:
+            conn.rollback()
+        raise
+    finally:
+        if conn:
+            conn.close()
+
+
+@contextmanager
+def get_db():
+    """Yield a connection and auto-commit/rollback."""
+    conn = _get_conn()
+    try:
+        yield conn
+        conn.commit()
+    except Exception:
+        conn.rollback()
+        raise
+    finally:
+        conn.close()

backend/main.py

@@ -1,16 +1,15 @@
 import os
-import re
+import uuid
 import tempfile
-import shutil
 import base64
 from typing import List, Dict, Any, Optional
 from contextlib import asynccontextmanager
+from datetime import datetime, timezone
 
-from fastapi import FastAPI, File, UploadFile, Form, HTTPException
+from fastapi import FastAPI, File, UploadFile, Form, HTTPException, Query
 from fastapi.middleware.cors import CORSMiddleware
-from fastapi.responses import JSONResponse, StreamingResponse
+from fastapi.responses import JSONResponse, StreamingResponse, FileResponse
-import torch
+from pydantic import BaseModel
-from transformers import AutoModel, AutoTokenizer
 from PIL import Image
 import uvicorn
 from decouple import config as env_config
@@ -24,51 +23,41 @@ from pdf_utils import (
     clean_markdown_content
 )
 from format_converter import DocumentConverter
+from database import init_db, get_db
+from providers import (
+    build_registry,
+    parse_detections,
+    clean_grounding_text,
+    ProviderError,
+    GROUNDING_MODES,
+)
+
+OCR_IMAGES_DIR = env_config("OCR_IMAGES_DIR", default="/data/ocr_images")
 
 # -----------------------------
-# Lifespan context for model loading
+# Lifespan context
 # -----------------------------
-model = None
+# The model registry holds all available OCR providers. Local models (e.g.
-tokenizer = None
+# DeepSeek-OCR) are loaded lazily on first use so an Ollama-only deployment
+# starts instantly and never touches the GPU.
+registry = None
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
-    """Load model on startup, cleanup on shutdown"""
+    """Build the model registry on startup."""
-    global model, tokenizer
+    global registry
 
-    # Environment setup
+    # Image storage directory
-    os.environ.pop("TRANSFORMERS_CACHE", None)
+    os.makedirs(OCR_IMAGES_DIR, exist_ok=True)
-    MODEL_NAME = env_config("MODEL_NAME", default="deepseek-ai/DeepSeek-OCR")
-    HF_HOME = env_config("HF_HOME", default="/models")
-    os.makedirs(HF_HOME, exist_ok=True)
 
-    # Load model
+    # Database
-    print(f"🚀 Loading {MODEL_NAME}...")
-    torch_dtype = torch.bfloat16
-    
-    tokenizer = AutoTokenizer.from_pretrained(
-        MODEL_NAME,
-        trust_remote_code=True,
-    )
-    
-    model = AutoModel.from_pretrained(
-        MODEL_NAME,
-        trust_remote_code=True,
-        use_safetensors=True,
-        attn_implementation="eager",
-        torch_dtype=torch_dtype,
-    ).eval().to("cuda")
-    
-    # Pad token setup
     try:
-        if getattr(tokenizer, "pad_token_id", None) is None and getattr(tokenizer, "eos_token_id", None) is not None:
+        init_db()
-            tokenizer.pad_token = tokenizer.eos_token
+    except Exception as exc:
-        if getattr(model.config, "pad_token_id", None) is None and getattr(tokenizer, "pad_token_id", None) is not None:
+        print(f"Warning: database initialization failed: {exc}")
-            model.config.pad_token_id = tokenizer.pad_token_id
-    except Exception:
-        pass
 
-    print("✅ Model loaded and ready!")
+    # OCR model registry (providers load their models lazily)
+    registry = build_registry()
 
     yield
 
@@ -97,155 +86,6 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
-# -----------------------------
-# Prompt builder
-# -----------------------------
-def build_prompt(
-    mode: str,
-    user_prompt: str,
-    grounding: bool,
-    find_term: Optional[str],
-    schema: Optional[str],
-    include_caption: bool,
-) -> str:
-    """Build the prompt based on mode"""
-    parts: List[str] = ["<image>"]
-    mode_requires_grounding = mode in {"find_ref", "layout_map", "pii_redact"}
-    if grounding or mode_requires_grounding:
-        parts.append("<|grounding|>")
-
-    instruction = ""
-    if mode == "plain_ocr":
-        instruction = "Free OCR."
-    elif mode == "markdown":
-        instruction = "Convert the document to markdown."
-    elif mode == "tables_csv":
-        instruction = (
-            "Extract every table and output CSV only. "
-            "Use commas, minimal quoting. If multiple tables, separate with a line containing '---'."
-        )
-    elif mode == "tables_md":
-        instruction = "Extract every table as GitHub-flavored Markdown tables. Output only the tables."
-    elif mode == "kv_json":
-        schema_text = schema.strip() if schema else "{}"
-        instruction = (
-            "Extract key fields and return strict JSON only. "
-            f"Use this schema (fill the values): {schema_text}"
-        )
-    elif mode == "figure_chart":
-        instruction = (
-            "Parse the figure. First extract any numeric series as a two-column table (x,y). "
-            "Then summarize the chart in 2 sentences. Output the table, then a line '---', then the summary."
-        )
-    elif mode == "find_ref":
-        key = (find_term or "").strip() or "Total"
-        instruction = f"Locate <|ref|>{key}<|/ref|> in the image."
-    elif mode == "layout_map":
-        instruction = (
-            'Return a JSON array of blocks with fields {"type":["title","paragraph","table","figure"],'
-            '"box":[x1,y1,x2,y2]}. Do not include any text content.'
-        )
-    elif mode == "pii_redact":
-        instruction = (
-            'Find all occurrences of emails, phone numbers, postal addresses, and IBANs. '
-            'Return a JSON array of objects {label, text, box:[x1,y1,x2,y2]}.'
-        )
-    elif mode == "multilingual":
-        instruction = "Free OCR. Detect the language automatically and output in the same script."
-    elif mode == "describe":
-        instruction = "Describe this image. Focus on visible key elements."
-    elif mode == "freeform":
-        instruction = user_prompt.strip() if user_prompt else "OCR this image."
-    else:
-        instruction = "OCR this image."
-
-    if include_caption and mode not in {"describe"}:
-        instruction = instruction + "\nThen add a one-paragraph description of the image."
-
-    parts.append(instruction)
-    return "\n".join(parts)
-
-# -----------------------------
-# Grounding parser
-# -----------------------------
-# Match a full detection block and capture the coordinates as the entire list expression
-# Examples of captured coords (including outer brackets):
-#  - [[312, 339, 480, 681]]
-#  - [[504, 700, 625, 910], [771, 570, 996, 996]]
-#  - [[110, 310, 255, 800], [312, 343, 479, 680], ...]
-# Using a greedy bracket capture ensures we include all inner lists up to the last ']' before </|det|>
-DET_BLOCK = re.compile(
-    r"<\|ref\|>(?P<label>.*?)<\|/ref\|>\s*<\|det\|>\s*(?P<coords>\[.*\])\s*<\|/det\|>",
-    re.DOTALL,
-)
-
-def clean_grounding_text(text: str) -> str:
-    """Remove grounding tags from text for display, keeping labels"""
-    # Replace <|ref|>label<|/ref|><|det|>[...any nested lists...]<|/det|> with just the label
-    cleaned = re.sub(
-        r"<\|ref\|>(.*?)<\|/ref\|>\s*<\|det\|>\s*\[.*\]\s*<\|/det\|>",
-        r"\1",
-        text,
-        flags=re.DOTALL,
-    )
-    # Also remove any standalone grounding tags
-    cleaned = re.sub(r"<\|grounding\|>", "", cleaned)
-    return cleaned.strip()
-
-def parse_detections(text: str, image_width: int, image_height: int) -> List[Dict[str, Any]]:
-    """Parse grounding boxes from text and scale from 0-999 normalized coords to actual image dimensions
-    
-    Handles both single and multiple bounding boxes:
-    - Single: <|ref|>label<|/ref|><|det|>[[x1,y1,x2,y2]]<|/det|>
-    - Multiple: <|ref|>label<|/ref|><|det|>[[x1,y1,x2,y2], [x1,y1,x2,y2], ...]<|/det|>
-    """
-    boxes: List[Dict[str, Any]] = []
-    for m in DET_BLOCK.finditer(text or ""):
-        label = m.group("label").strip()
-        coords_str = m.group("coords").strip()
-
-        print(f"🔍 DEBUG: Found detection for '{label}'")
-        print(f"📦 Raw coords string (with brackets): {coords_str}")
-
-        try:
-            import ast
-
-            # Parse the full bracket expression directly (handles single and multiple)
-            parsed = ast.literal_eval(coords_str)
-
-            # Normalize to a list of lists
-            if (
-                isinstance(parsed, list)
-                and len(parsed) == 4
-                and all(isinstance(n, (int, float)) for n in parsed)
-            ):
-                # Single box provided as [x1,y1,x2,y2]
-                box_coords = [parsed]
-                print("📦 Single box (flat list) detected")
-            elif isinstance(parsed, list):
-                box_coords = parsed
-                print(f"📦 Boxes detected: {len(box_coords)}")
-            else:
-                raise ValueError("Unsupported coords structure")
-
-            # Process each box
-            for idx, box in enumerate(box_coords):
-                if isinstance(box, (list, tuple)) and len(box) >= 4:
-                    x1 = int(float(box[0]) / 999 * image_width)
-                    y1 = int(float(box[1]) / 999 * image_height)
-                    x2 = int(float(box[2]) / 999 * image_width)
-                    y2 = int(float(box[3]) / 999 * image_height)
-                    print(f"  Box {idx+1}: {box} → [{x1}, {y1}, {x2}, {y2}]")
-                    boxes.append({"label": label, "box": [x1, y1, x2, y2]})
-                else:
-                    print(f"  ⚠️ Skipping invalid box: {box}")
-        except Exception as e:
-            print(f"❌ Parsing failed: {e}")
-            continue
-    
-    print(f"🎯 Total boxes parsed: {len(boxes)}")
-    return boxes
-
 # -----------------------------
 # Routes
 # -----------------------------
@@ -255,11 +95,38 @@ async def root():
 
 @app.get("/health")
 async def health():
-    return {"status": "healthy", "model_loaded": model is not None}
+    return {"status": "healthy", "models": registry.list_models() if registry else []}
+
+
+@app.get("/api/models")
+async def list_models():
+    """List the OCR models available for selection in the UI."""
+    if registry is None:
+        raise HTTPException(status_code=503, detail="Model registry not ready.")
+    return JSONResponse({"models": registry.list_models()})
+
+
+def _resolve_provider(model_id: Optional[str], mode: str):
+    """Look up the provider and reject capability mismatches (e.g. grounding)."""
+    if registry is None:
+        raise HTTPException(status_code=503, detail="Model registry not ready.")
+    try:
+        provider = registry.get(model_id)
+    except ProviderError as exc:
+        raise HTTPException(status_code=400, detail=str(exc))
+
+    if mode in GROUNDING_MODES and not provider.capabilities.get("grounding"):
+        raise HTTPException(
+            status_code=400,
+            detail=f"Model '{provider.label}' does not support grounding modes (e.g. {mode}).",
+        )
+    return provider
+
 
 @app.post("/api/ocr")
 async def ocr_inference(
     image: UploadFile = File(...),
+    model: Optional[str] = Form(None),
     mode: str = Form("plain_ocr"),
     prompt: str = Form(""),
     grounding: bool = Form(False),
@@ -275,32 +142,18 @@ async def ocr_inference(
     Perform OCR inference on uploaded image
 
     - **image**: Image file to process
+    - **model**: OCR model id (see GET /api/models); defaults to the registry default
     - **mode**: OCR mode (plain_ocr, markdown, tables_csv, etc.)
     - **prompt**: Custom prompt for freeform mode
-    - **grounding**: Enable grounding boxes
+    - **grounding**: Enable grounding boxes (DeepSeek only)
     - **include_caption**: Add image description
     - **find_term**: Term to find (for find_ref mode)
     - **schema**: JSON schema (for kv_json mode)
-    - **base_size**: Base processing size
+    - **base_size/image_size/crop_mode/test_compress**: DeepSeek processing options
-    - **image_size**: Image size parameter
-    - **crop_mode**: Enable crop mode
-    - **test_compress**: Test compression
     """
-    if model is None or tokenizer is None:
+    provider = _resolve_provider(model, mode)
-        raise HTTPException(status_code=503, detail="Model not loaded yet")
-    
-    # Build prompt
-    prompt_text = build_prompt(
-        mode=mode,
-        user_prompt=prompt,
-        grounding=grounding,
-        find_term=find_term,
-        schema=schema,
-        include_caption=include_caption,
-    )
 
     tmp_img = None
-    out_dir = None
     try:
         # Save uploaded file
         with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as tmp:
@@ -315,42 +168,27 @@ async def ocr_inference(
         except Exception:
             orig_w = orig_h = None
 
-        out_dir = tempfile.mkdtemp(prefix="dsocr_")
+        # Run inference through the selected provider
-        
+        text = provider.run(
-        # Run inference
+            tmp_img,
-        res = model.infer(
+            mode=mode,
-            tokenizer,
+            prompt=prompt,
-            prompt=prompt_text,
+            grounding=grounding,
-            image_file=tmp_img,
+            find_term=find_term,
-            output_path=out_dir,
+            schema=schema,
-            base_size=base_size,
+            include_caption=include_caption,
-            image_size=image_size,
+            options={
-            crop_mode=crop_mode,
+                "base_size": base_size,
-            save_results=False,
+                "image_size": image_size,
-            test_compress=test_compress,
+                "crop_mode": crop_mode,
-            eval_mode=True,
+                "test_compress": test_compress,
+            },
         )
 
-        # Normalize response
-        if isinstance(res, str):
-            text = res.strip()
-        elif isinstance(res, dict) and "text" in res:
-            text = str(res["text"]).strip()
-        elif isinstance(res, (list, tuple)):
-            text = "\n".join(map(str, res)).strip()
-        else:
-            text = ""
-        
-        # Fallback: check output file
-        if not text:
-            mmd = os.path.join(out_dir, "result.mmd")
-            if os.path.exists(mmd):
-                with open(mmd, "r", encoding="utf-8") as fh:
-                    text = fh.read().strip()
         if not text:
             text = "No text returned by model."
 
-        # Parse grounding boxes with proper coordinate scaling
+        # Parse grounding boxes (no-op for providers/text without grounding tokens)
         boxes = parse_detections(text, orig_w or 1, orig_h or 1) if ("<|det|>" in text or "<|ref|>" in text) else []
 
         # Clean grounding tags from display text, but keep the labels
@@ -367,14 +205,21 @@ async def ocr_inference(
             "boxes": boxes,
             "image_dims": {"w": orig_w, "h": orig_h},
             "metadata": {
+                "model": provider.id,
+                "model_label": provider.label,
                 "mode": mode,
-                "grounding": grounding or (mode in {"find_ref","layout_map","pii_redact"}),
+                "grounding": grounding or (mode in GROUNDING_MODES),
                 "base_size": base_size,
                 "image_size": image_size,
                 "crop_mode": crop_mode
             }
         })
 
+    except ProviderError as e:
+        print(f"OCR provider error: {e}")
+        raise HTTPException(status_code=502, detail=str(e))
+    except HTTPException:
+        raise
     except Exception as e:
         print(f"OCR inference error: {type(e).__name__}: {str(e)}")
         raise HTTPException(status_code=500, detail="An internal error occurred during OCR processing.")
@@ -385,12 +230,11 @@ async def ocr_inference(
                 os.remove(tmp_img)
             except Exception:
                 pass
-        if out_dir:
-            shutil.rmtree(out_dir, ignore_errors=True)
 
 @app.post("/api/process-pdf")
 async def process_pdf(
     pdf_file: UploadFile = File(...),
+    model: Optional[str] = Form(None),
     mode: str = Form("plain_ocr"),
     prompt: str = Form(""),
     output_format: str = Form("markdown"),  # markdown, html, docx, json
@@ -417,8 +261,7 @@ async def process_pdf(
     - **image_size**: Image size parameter
     - **crop_mode**: Enable crop mode
     """
-    if model is None or tokenizer is None:
+    provider = _resolve_provider(model, mode)
-        raise HTTPException(status_code=503, detail="Model not loaded yet")
 
     # Validate output format
     if output_format not in ["markdown", "html", "docx", "json"]:
@@ -441,56 +284,32 @@ async def process_pdf(
         for page_idx, img in enumerate(images):
             print(f"🔍 Processing page {page_idx + 1}/{total_pages}...")
 
-            # Build prompt for this page
-            prompt_text = build_prompt(
-                mode=mode,
-                user_prompt=prompt,
-                grounding=grounding,
-                find_term=None,
-                schema=None,
-                include_caption=include_caption,
-            )
-
             # Save image temporarily
             tmp_img = None
-            out_dir = None
             try:
                 with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as tmp:
                     img.save(tmp, format="PNG")
                     tmp_img = tmp.name
 
                 orig_w, orig_h = img.size
-                out_dir = tempfile.mkdtemp(prefix="dsocr_pdf_")
 
-                # Run inference
+                # Run inference through the selected provider
-                res = model.infer(
+                text = provider.run(
-                    tokenizer,
+                    tmp_img,
-                    prompt=prompt_text,
+                    mode=mode,
-                    image_file=tmp_img,
+                    prompt=prompt,
-                    output_path=out_dir,
+                    grounding=grounding,
-                    base_size=base_size,
+                    find_term=None,
-                    image_size=image_size,
+                    schema=None,
-                    crop_mode=crop_mode,
+                    include_caption=include_caption,
-                    save_results=False,
+                    options={
-                    test_compress=False,
+                        "base_size": base_size,
-                    eval_mode=True,
+                        "image_size": image_size,
+                        "crop_mode": crop_mode,
+                        "test_compress": False,
+                    },
                 )
 
-                # Normalize response
-                if isinstance(res, str):
-                    text = res.strip()
-                elif isinstance(res, dict) and "text" in res:
-                    text = str(res["text"]).strip()
-                elif isinstance(res, (list, tuple)):
-                    text = "\n".join(map(str, res)).strip()
-                else:
-                    text = ""
-
-                if not text:
-                    mmd = os.path.join(out_dir, "result.mmd")
-                    if os.path.exists(mmd):
-                        with open(mmd, "r", encoding="utf-8") as fh:
-                            text = fh.read().strip()
                 if not text:
                     text = f"No text returned for page {page_idx + 1}."
 
@@ -535,8 +354,6 @@ async def process_pdf(
                         os.remove(tmp_img)
                     except Exception:
                         pass
-                if out_dir:
-                    shutil.rmtree(out_dir, ignore_errors=True)
 
         print(f"✅ Processed all {total_pages} pages")
 
@@ -547,6 +364,8 @@ async def process_pdf(
                 "total_pages": total_pages,
                 "pages": pages_content,
                 "metadata": {
+                    "model": provider.id,
+                    "model_label": provider.label,
                     "mode": mode,
                     "grounding": grounding,
                     "extract_images": extract_images,
@@ -575,12 +394,468 @@ async def process_pdf(
                 headers={"Content-Disposition": f"attachment; filename=ocr_result.docx"}
             )
 
+    except ProviderError as e:
+        print(f"PDF provider error: {e}")
+        raise HTTPException(status_code=502, detail=str(e))
     except Exception as e:
         import traceback
         print(f"Error processing PDF: {e}")
         print(traceback.format_exc())
         raise HTTPException(status_code=500, detail="An internal error occurred during PDF processing.")
 
+# -----------------------------
+# Job management routes
+# -----------------------------
+
+class ReviewRequest(BaseModel):
+    reviewed_text: str
+    reviewer_name: str
+    author: Optional[str] = None
+    book: Optional[str] = None
+    chapter: Optional[str] = None
+    page: Optional[str] = None
+    describe_text: Optional[str] = None
+    freeform_text: Optional[str] = None
+
+
+def _job_row_to_dict(row) -> Dict[str, Any]:
+    """Convert a DB row (RealDictRow) to a plain dict with serialisable values."""
+    d = dict(row)
+    for key, val in d.items():
+        if isinstance(val, datetime):
+            d[key] = val.isoformat()
+        elif val is not None and hasattr(val, '__str__') and type(val).__name__ == 'UUID':
+            d[key] = str(val)
+    return d
+
+
+@app.post("/api/jobs")
+async def commit_job(
+    image: UploadFile = File(...),
+    author: str = Form(""),
+    book: str = Form(""),
+    chapter: str = Form(""),
+    page: str = Form(""),
+    ocr_text: str = Form(""),
+    describe_text: str = Form(""),
+    freeform_text: str = Form(""),
+    mode: str = Form("plain_ocr"),
+    ocr_model: str = Form(""),
+):
+    """Commit an OCR job: save the image and insert a DB record."""
+    job_id = str(uuid.uuid4())
+
+    # Determine file extension from original filename or content type
+    original_filename = image.filename or "image"
+    ext = os.path.splitext(original_filename)[1].lower()
+    if not ext:
+        ct = (image.content_type or "").lower()
+        ext_map = {
+            "image/png": ".png", "image/jpeg": ".jpg", "image/jpg": ".jpg",
+            "image/webp": ".webp", "image/gif": ".gif", "image/bmp": ".bmp",
+        }
+        ext = ext_map.get(ct, ".png")
+
+    image_path = os.path.join(OCR_IMAGES_DIR, f"{job_id}{ext}")
+
+    try:
+        content = await image.read()
+        with open(image_path, "wb") as f:
+            f.write(content)
+    except Exception as exc:
+        raise HTTPException(status_code=500, detail="Failed to save image file.")
+
+    try:
+        with get_db() as conn:
+            with conn.cursor() as cur:
+                cur.execute(
+                    """
+                    INSERT INTO ocr_jobs
+                        (id, author, book, chapter, page, image_path, original_filename,
+                         ocr_text, describe_text, freeform_text, mode, ocr_model, status)
+                    VALUES (%s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, 'unreviewed')
+                    RETURNING *
+                    """,
+                    (job_id, author or None, book or None, chapter or None,
+                     page or None, image_path, original_filename,
+                     ocr_text or None, describe_text or None, freeform_text or None,
+                     mode, ocr_model or None),
+                )
+                row = cur.fetchone()
+    except Exception as exc:
+        # Clean up saved image if DB insert fails
+        try:
+            os.remove(image_path)
+        except Exception:
+            pass
+        # Unique constraint violation (author + chapter + page already exists)
+        if getattr(exc, 'pgcode', None) == '23505':
+            raise HTTPException(
+                status_code=409,
+                detail="A job with this Author, Chapter, and Page already exists."
+            )
+        print(f"Job commit DB error: {exc}")
+        raise HTTPException(status_code=500, detail="Failed to save job to database.")
+
+    return JSONResponse(_job_row_to_dict(row), status_code=201)
+
+
+@app.get("/api/jobs")
+async def list_jobs(
+    search: Optional[str] = Query(None, description="General text search across all fields"),
+    author: Optional[str] = Query(None),
+    book: Optional[str] = Query(None),
+    chapter: Optional[str] = Query(None),
+    status: Optional[str] = Query(None, description="unreviewed | reviewed"),
+    limit: int = Query(20, ge=1, le=200),
+    offset: int = Query(0, ge=0),
+):
+    """Search and list jobs. All filters are optional and combinable."""
+    conditions = []
+    params: List[Any] = []
+
+    if search:
+        conditions.append(
+            "(author ILIKE %s OR book ILIKE %s OR chapter ILIKE %s "
+            "OR page ILIKE %s OR ocr_text ILIKE %s OR reviewer_name ILIKE %s)"
+        )
+        like = f"%{search}%"
+        params.extend([like, like, like, like, like, like])
+
+    if author:
+        conditions.append("author ILIKE %s")
+        params.append(f"%{author}%")
+
+    if book:
+        conditions.append("book ILIKE %s")
+        params.append(f"%{book}%")
+
+    if chapter:
+        conditions.append("chapter ILIKE %s")
+        params.append(f"%{chapter}%")
+
+    if status:
+        conditions.append("status = %s")
+        params.append(status)
+
+    where = ("WHERE " + " AND ".join(conditions)) if conditions else ""
+
+    try:
+        with get_db() as conn:
+            with conn.cursor() as cur:
+                cur.execute(
+                    f"SELECT COUNT(*) AS total FROM ocr_jobs {where}",
+                    params,
+                )
+                total = cur.fetchone()["total"]
+
+                cur.execute(
+                    f"""
+                    SELECT id, author, book, chapter, page, submitted_at, status,
+                           reviewer_name, reviewed_at, mode, ocr_model, original_filename
+                    FROM ocr_jobs {where}
+                    ORDER BY submitted_at DESC
+                    LIMIT %s OFFSET %s
+                    """,
+                    params + [limit, offset],
+                )
+                rows = [_job_row_to_dict(r) for r in cur.fetchall()]
+    except Exception as exc:
+        print(f"list_jobs DB error: {exc}")
+        raise HTTPException(status_code=500, detail="Database error.")
+
+    return JSONResponse({"total": total, "limit": limit, "offset": offset, "jobs": rows})
+
+
+@app.get("/api/jobs/suggestions")
+async def job_suggestions():
+    """Return distinct values for author, book, chapter, and reviewer_name to power autocomplete."""
+    try:
+        with get_db() as conn:
+            with conn.cursor() as cur:
+                cur.execute("""
+                    SELECT
+                        array_remove(array_agg(DISTINCT author ORDER BY author), NULL) AS authors,
+                        array_remove(array_agg(DISTINCT book ORDER BY book), NULL) AS books,
+                        array_remove(array_agg(DISTINCT chapter ORDER BY chapter), NULL) AS chapters,
+                        array_remove(array_agg(DISTINCT reviewer_name ORDER BY reviewer_name), NULL) AS reviewers
+                    FROM ocr_jobs
+                """)
+                row = cur.fetchone()
+    except Exception as exc:
+        print(f"suggestions DB error: {exc}")
+        raise HTTPException(status_code=500, detail="Database error.")
+
+    return JSONResponse({
+        "authors": row["authors"] or [],
+        "books": row["books"] or [],
+        "chapters": row["chapters"] or [],
+        "reviewers": row["reviewers"] or [],
+    })
+
+
+@app.get("/api/jobs/{job_id}")
+async def get_job(job_id: str):
+    """Retrieve full job record including OCR text."""
+    try:
+        uuid.UUID(job_id)
+    except ValueError:
+        raise HTTPException(status_code=400, detail="Invalid job ID.")
+
+    try:
+        with get_db() as conn:
+            with conn.cursor() as cur:
+                cur.execute("SELECT * FROM ocr_jobs WHERE id = %s", (job_id,))
+                row = cur.fetchone()
+    except Exception as exc:
+        print(f"get_job DB error: {exc}")
+        raise HTTPException(status_code=500, detail="Database error.")
+
+    if not row:
+        raise HTTPException(status_code=404, detail="Job not found.")
+
+    return JSONResponse(_job_row_to_dict(row))
+
+
+@app.get("/api/jobs/{job_id}/image")
+async def get_job_image(job_id: str):
+    """Serve the stored image for a job."""
+    try:
+        uuid.UUID(job_id)
+    except ValueError:
+        raise HTTPException(status_code=400, detail="Invalid job ID.")
+
+    try:
+        with get_db() as conn:
+            with conn.cursor() as cur:
+                cur.execute("SELECT image_path FROM ocr_jobs WHERE id = %s", (job_id,))
+                row = cur.fetchone()
+    except Exception as exc:
+        print(f"get_job_image DB error: {exc}")
+        raise HTTPException(status_code=500, detail="Database error.")
+
+    if not row:
+        raise HTTPException(status_code=404, detail="Job not found.")
+
+    path = row["image_path"]
+    if not os.path.isfile(path):
+        raise HTTPException(status_code=404, detail="Image file not found on disk.")
+
+    return FileResponse(path)
+
+
+@app.put("/api/jobs/{job_id}/review")
+async def review_job(job_id: str, body: ReviewRequest):
+    """Mark a job as reviewed with the corrected text and reviewer name."""
+    try:
+        uuid.UUID(job_id)
+    except ValueError:
+        raise HTTPException(status_code=400, detail="Invalid job ID.")
+
+    try:
+        with get_db() as conn:
+            with conn.cursor() as cur:
+                cur.execute(
+                    """
+                    UPDATE ocr_jobs
+                    SET status = 'reviewed',
+                        reviewed_text = %s,
+                        reviewer_name = %s,
+                        reviewed_at = NOW(),
+                        author = %s,
+                        book = %s,
+                        chapter = %s,
+                        page = %s,
+                        describe_text = %s,
+                        freeform_text = %s
+                    WHERE id = %s
+                    RETURNING *
+                    """,
+                    (
+                        body.reviewed_text,
+                        body.reviewer_name,
+                        body.author or None,
+                        body.book or None,
+                        body.chapter or None,
+                        body.page or None,
+                        body.describe_text,
+                        body.freeform_text,
+                        job_id,
+                    ),
+                )
+                row = cur.fetchone()
+    except Exception as exc:
+        print(f"review_job DB error: {exc}")
+        raise HTTPException(status_code=500, detail="Database error.")
+
+    if not row:
+        raise HTTPException(status_code=404, detail="Job not found.")
+
+    return JSONResponse(_job_row_to_dict(row))
+
+
+class StatusRequest(BaseModel):
+    status: str
+    reviewer_name: Optional[str] = None
+
+
+@app.put("/api/jobs/{job_id}/status")
+async def set_job_status(job_id: str, body: StatusRequest):
+    """Toggle a job's reviewed status without touching its text or metadata.
+
+    Marking 'reviewed' requires a reviewer_name and stamps reviewed_at.
+    Marking 'unreviewed' clears reviewed_at while preserving reviewed_text.
+    """
+    try:
+        uuid.UUID(job_id)
+    except ValueError:
+        raise HTTPException(status_code=400, detail="Invalid job ID.")
+
+    if body.status not in ("reviewed", "unreviewed"):
+        raise HTTPException(status_code=400, detail="status must be 'reviewed' or 'unreviewed'.")
+
+    if body.status == "reviewed" and not (body.reviewer_name or "").strip():
+        raise HTTPException(status_code=400, detail="Reviewer name is required to mark reviewed.")
+
+    try:
+        with get_db() as conn:
+            with conn.cursor() as cur:
+                if body.status == "reviewed":
+                    cur.execute(
+                        """
+                        UPDATE ocr_jobs
+                        SET status = 'reviewed',
+                            reviewer_name = %s,
+                            reviewed_at = NOW()
+                        WHERE id = %s
+                        RETURNING *
+                        """,
+                        (body.reviewer_name.strip(), job_id),
+                    )
+                else:
+                    cur.execute(
+                        """
+                        UPDATE ocr_jobs
+                        SET status = 'unreviewed',
+                            reviewed_at = NULL
+                        WHERE id = %s
+                        RETURNING *
+                        """,
+                        (job_id,),
+                    )
+                row = cur.fetchone()
+    except Exception as exc:
+        print(f"set_job_status DB error: {exc}")
+        raise HTTPException(status_code=500, detail="Database error.")
+
+    if not row:
+        raise HTTPException(status_code=404, detail="Job not found.")
+
+    return JSONResponse(_job_row_to_dict(row))
+
+
+class JobDescribeRequest(BaseModel):
+    model: Optional[str] = None
+
+
+@app.post("/api/jobs/{job_id}/describe")
+async def describe_job(job_id: str, body: JobDescribeRequest):
+    """Run Describe mode on a job's stored image and save the result to describe_text."""
+    try:
+        uuid.UUID(job_id)
+    except ValueError:
+        raise HTTPException(status_code=400, detail="Invalid job ID.")
+
+    # Look up the stored image for this job
+    try:
+        with get_db() as conn:
+            with conn.cursor() as cur:
+                cur.execute("SELECT image_path FROM ocr_jobs WHERE id = %s", (job_id,))
+                row = cur.fetchone()
+    except Exception as exc:
+        print(f"describe_job lookup DB error: {exc}")
+        raise HTTPException(status_code=500, detail="Database error.")
+
+    if not row:
+        raise HTTPException(status_code=404, detail="Job not found.")
+    image_path = row["image_path"]
+    if not image_path or not os.path.isfile(image_path):
+        raise HTTPException(status_code=404, detail="Image file not found on disk.")
+
+    provider = _resolve_provider(body.model, "describe")
+
+    try:
+        text = provider.run(
+            image_path,
+            mode="describe",
+            prompt="",
+            grounding=False,
+            find_term=None,
+            schema=None,
+            include_caption=False,
+            options={"base_size": 1024, "image_size": 640, "crop_mode": True, "test_compress": False},
+        )
+    except ProviderError as e:
+        print(f"describe_job provider error: {e}")
+        raise HTTPException(status_code=502, detail=str(e))
+    except Exception as e:
+        print(f"describe_job inference error: {type(e).__name__}: {e}")
+        raise HTTPException(status_code=500, detail="An internal error occurred during description.")
+
+    display_text = clean_grounding_text(text) if ("<|ref|>" in text or "<|grounding|>" in text) else text
+
+    # Persist the generated description on the job
+    try:
+        with get_db() as conn:
+            with conn.cursor() as cur:
+                cur.execute(
+                    "UPDATE ocr_jobs SET describe_text = %s WHERE id = %s RETURNING *",
+                    (display_text, job_id),
+                )
+                updated = cur.fetchone()
+    except Exception as exc:
+        print(f"describe_job save DB error: {exc}")
+        raise HTTPException(status_code=500, detail="Database error.")
+
+    if not updated:
+        raise HTTPException(status_code=404, detail="Job not found.")
+
+    return JSONResponse(_job_row_to_dict(updated))
+
+
+@app.delete("/api/jobs/{job_id}")
+async def delete_job(job_id: str):
+    """Delete a job record and its stored image."""
+    try:
+        uuid.UUID(job_id)
+    except ValueError:
+        raise HTTPException(status_code=400, detail="Invalid job ID.")
+
+    try:
+        with get_db() as conn:
+            with conn.cursor() as cur:
+                cur.execute(
+                    "DELETE FROM ocr_jobs WHERE id = %s RETURNING image_path",
+                    (job_id,),
+                )
+                row = cur.fetchone()
+    except Exception as exc:
+        print(f"delete_job DB error: {exc}")
+        raise HTTPException(status_code=500, detail="Database error.")
+
+    if not row:
+        raise HTTPException(status_code=404, detail="Job not found.")
+
+    # Best-effort removal of the stored image file
+    try:
+        if row["image_path"] and os.path.isfile(row["image_path"]):
+            os.remove(row["image_path"])
+    except Exception:
+        pass
+
+    return JSONResponse({"deleted": job_id})
+
+
 if __name__ == "__main__":
     host = env_config("API_HOST", default="0.0.0.0")
     port = env_config("API_PORT", default=8000, cast=int)

backend/providers.py

@@ -0,0 +1,489 @@
+"""
+OCR provider abstraction.
+
+Each provider knows how to turn an image + a semantic OCR request (mode, prompt,
+options) into raw model text. DeepSeek-specific prompt tokens and grounding-box
+parsing live here too so the FastAPI routes stay model-agnostic.
+
+Two providers ship today:
+  - DeepSeekLocalProvider  -> the local HF transformers DeepSeek-OCR model (GPU)
+  - OllamaProvider         -> any vision model served by an external Ollama host
+
+The registry is built from environment variables at startup (see build_registry()).
+"""
+
+import os
+import re
+import base64
+import tempfile
+import shutil
+from abc import ABC, abstractmethod
+from typing import List, Dict, Any, Optional
+
+from decouple import config as env_config
+
+# httpx is only needed when an Ollama model is actually used; import lazily so the
+# backend can run DeepSeek-only without the dependency installed.
+try:
+    import httpx
+except Exception:  # pragma: no cover - exercised only when httpx is missing
+    httpx = None
+
+
+# =============================================================================
+# Prompt builders
+# =============================================================================
+def build_prompt(
+    mode: str,
+    user_prompt: str,
+    grounding: bool,
+    find_term: Optional[str],
+    schema: Optional[str],
+    include_caption: bool,
+) -> str:
+    """Build the DeepSeek-OCR prompt (with its special tokens) based on mode."""
+    parts: List[str] = ["<image>"]
+    mode_requires_grounding = mode in {"find_ref", "layout_map", "pii_redact"}
+    if grounding or mode_requires_grounding:
+        parts.append("<|grounding|>")
+
+    parts.append(_instruction_for_mode(mode, user_prompt, find_term, schema, include_caption))
+    return "\n".join(parts)
+
+
+def build_ollama_prompt(
+    mode: str,
+    user_prompt: str,
+    find_term: Optional[str],
+    schema: Optional[str],
+    include_caption: bool,
+) -> str:
+    """Build a plain natural-language prompt for a generic vision model.
+
+    No DeepSeek grounding tokens — Ollama vision models receive the image
+    separately and respond in plain text.
+    """
+    if mode == "plain_ocr":
+        instruction = (
+            "Transcribe all of the text in this image exactly as it appears, "
+            "preserving line breaks and reading order. Output only the transcribed "
+            "text with no commentary."
+        )
+    elif mode == "markdown":
+        instruction = (
+            "Convert this document image to clean GitHub-flavored Markdown, "
+            "preserving headings, lists, and tables. Output only the Markdown."
+        )
+    elif mode == "tables_csv":
+        instruction = (
+            "Extract every table in this image and output CSV only. Use commas with "
+            "minimal quoting. If there are multiple tables, separate them with a line "
+            "containing '---'. Output only the CSV."
+        )
+    elif mode == "tables_md":
+        instruction = (
+            "Extract every table in this image as GitHub-flavored Markdown tables. "
+            "Output only the tables."
+        )
+    elif mode == "kv_json":
+        schema_text = schema.strip() if schema else "{}"
+        instruction = (
+            "Extract the key fields from this image and return strict JSON only "
+            f"(no prose). Use this schema, filling in the values: {schema_text}"
+        )
+    elif mode == "figure_chart":
+        instruction = (
+            "Parse the figure in this image. First extract any numeric series as a "
+            "two-column table (x,y). Then add a line containing '---' followed by a "
+            "two-sentence summary of the chart."
+        )
+    elif mode == "find_ref":
+        key = (find_term or "").strip() or "Total"
+        instruction = (
+            f"Find every occurrence of '{key}' in this image and quote the surrounding "
+            "text for each match. If it does not appear, say so."
+        )
+    elif mode == "layout_map":
+        instruction = (
+            'Identify the layout blocks in this image and return a JSON array of '
+            'objects {"type": one of ["title","paragraph","table","figure"]}. '
+            "Do not include the text content."
+        )
+    elif mode == "pii_redact":
+        instruction = (
+            "Find all emails, phone numbers, postal addresses, and IBANs in this image. "
+            'Return a JSON array of objects {"label", "text"}.'
+        )
+    elif mode == "multilingual":
+        instruction = (
+            "Transcribe all of the text in this image exactly, detecting the language "
+            "automatically and preserving the original script. Output only the text."
+        )
+    elif mode == "describe":
+        instruction = "Describe this image, focusing on the key visible elements."
+    elif mode == "freeform":
+        instruction = user_prompt.strip() if user_prompt else "Transcribe the text in this image."
+    else:
+        instruction = "Transcribe the text in this image."
+
+    if include_caption and mode != "describe":
+        instruction += "\nThen add a one-paragraph description of the image."
+
+    return instruction
+
+
+def _instruction_for_mode(
+    mode: str,
+    user_prompt: str,
+    find_term: Optional[str],
+    schema: Optional[str],
+    include_caption: bool,
+) -> str:
+    """The DeepSeek instruction text (without the <image>/<|grounding|> prefix tokens)."""
+    if mode == "plain_ocr":
+        instruction = "Free OCR."
+    elif mode == "markdown":
+        instruction = "Convert the document to markdown."
+    elif mode == "tables_csv":
+        instruction = (
+            "Extract every table and output CSV only. "
+            "Use commas, minimal quoting. If multiple tables, separate with a line containing '---'."
+        )
+    elif mode == "tables_md":
+        instruction = "Extract every table as GitHub-flavored Markdown tables. Output only the tables."
+    elif mode == "kv_json":
+        schema_text = schema.strip() if schema else "{}"
+        instruction = (
+            "Extract key fields and return strict JSON only. "
+            f"Use this schema (fill the values): {schema_text}"
+        )
+    elif mode == "figure_chart":
+        instruction = (
+            "Parse the figure. First extract any numeric series as a two-column table (x,y). "
+            "Then summarize the chart in 2 sentences. Output the table, then a line '---', then the summary."
+        )
+    elif mode == "find_ref":
+        key = (find_term or "").strip() or "Total"
+        instruction = f"Locate <|ref|>{key}<|/ref|> in the image."
+    elif mode == "layout_map":
+        instruction = (
+            'Return a JSON array of blocks with fields {"type":["title","paragraph","table","figure"],'
+            '"box":[x1,y1,x2,y2]}. Do not include any text content.'
+        )
+    elif mode == "pii_redact":
+        instruction = (
+            'Find all occurrences of emails, phone numbers, postal addresses, and IBANs. '
+            'Return a JSON array of objects {label, text, box:[x1,y1,x2,y2]}.'
+        )
+    elif mode == "multilingual":
+        instruction = "Free OCR. Detect the language automatically and output in the same script."
+    elif mode == "describe":
+        instruction = "Describe this image. Focus on visible key elements."
+    elif mode == "freeform":
+        instruction = user_prompt.strip() if user_prompt else "OCR this image."
+    else:
+        instruction = "OCR this image."
+
+    if include_caption and mode != "describe":
+        instruction = instruction + "\nThen add a one-paragraph description of the image."
+
+    return instruction
+
+
+# =============================================================================
+# Grounding parser (DeepSeek-specific; no-op on plain text)
+# =============================================================================
+DET_BLOCK = re.compile(
+    r"<\|ref\|>(?P<label>.*?)<\|/ref\|>\s*<\|det\|>\s*(?P<coords>\[.*\])\s*<\|/det\|>",
+    re.DOTALL,
+)
+
+
+def clean_grounding_text(text: str) -> str:
+    """Remove grounding tags from text for display, keeping labels."""
+    cleaned = re.sub(
+        r"<\|ref\|>(.*?)<\|/ref\|>\s*<\|det\|>\s*\[.*\]\s*<\|/det\|>",
+        r"\1",
+        text,
+        flags=re.DOTALL,
+    )
+    cleaned = re.sub(r"<\|grounding\|>", "", cleaned)
+    return cleaned.strip()
+
+
+def parse_detections(text: str, image_width: int, image_height: int) -> List[Dict[str, Any]]:
+    """Parse grounding boxes from text and scale 0-999 normalized coords to pixels."""
+    boxes: List[Dict[str, Any]] = []
+    for m in DET_BLOCK.finditer(text or ""):
+        label = m.group("label").strip()
+        coords_str = m.group("coords").strip()
+
+        try:
+            import ast
+
+            parsed = ast.literal_eval(coords_str)
+
+            if (
+                isinstance(parsed, list)
+                and len(parsed) == 4
+                and all(isinstance(n, (int, float)) for n in parsed)
+            ):
+                box_coords = [parsed]
+            elif isinstance(parsed, list):
+                box_coords = parsed
+            else:
+                raise ValueError("Unsupported coords structure")
+
+            for box in box_coords:
+                if isinstance(box, (list, tuple)) and len(box) >= 4:
+                    x1 = int(float(box[0]) / 999 * image_width)
+                    y1 = int(float(box[1]) / 999 * image_height)
+                    x2 = int(float(box[2]) / 999 * image_width)
+                    y2 = int(float(box[3]) / 999 * image_height)
+                    boxes.append({"label": label, "box": [x1, y1, x2, y2]})
+        except Exception as e:
+            print(f"❌ Grounding parse failed: {e}")
+            continue
+
+    return boxes
+
+
+# =============================================================================
+# Providers
+# =============================================================================
+GROUNDING_MODES = {"find_ref", "layout_map", "pii_redact"}
+
+
+class ProviderError(Exception):
+    """Raised when a provider cannot fulfil a request (e.g. backend unreachable)."""
+
+
+class OCRProvider(ABC):
+    """Turns an image + OCR request into raw model text."""
+
+    id: str
+    label: str
+    capabilities: Dict[str, Any]
+
+    @abstractmethod
+    def run(
+        self,
+        image_path: str,
+        *,
+        mode: str,
+        prompt: str,
+        grounding: bool,
+        find_term: Optional[str],
+        schema: Optional[str],
+        include_caption: bool,
+        options: Dict[str, Any],
+    ) -> str:
+        """Return the raw text output of the model for this image/request."""
+
+    def info(self) -> Dict[str, Any]:
+        return {"id": self.id, "label": self.label, "capabilities": self.capabilities}
+
+
+class DeepSeekLocalProvider(OCRProvider):
+    """Local HF transformers DeepSeek-OCR model. Loaded lazily on first use."""
+
+    def __init__(self):
+        self.id = "deepseek-local"
+        self.label = "DeepSeek-OCR (local GPU)"
+        self.capabilities = {"grounding": True, "advanced_settings": True}
+        self._model = None
+        self._tokenizer = None
+
+    @property
+    def loaded(self) -> bool:
+        return self._model is not None and self._tokenizer is not None
+
+    def _ensure_loaded(self):
+        if self.loaded:
+            return
+
+        # Heavy imports kept local so an Ollama-only deployment never needs torch.
+        import torch
+        from transformers import AutoModel, AutoTokenizer
+
+        os.environ.pop("TRANSFORMERS_CACHE", None)
+        model_name = env_config("MODEL_NAME", default="deepseek-ai/DeepSeek-OCR")
+        hf_home = env_config("HF_HOME", default="/models")
+        os.makedirs(hf_home, exist_ok=True)
+
+        print(f"🚀 Loading {model_name}...")
+        tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
+        model = AutoModel.from_pretrained(
+            model_name,
+            trust_remote_code=True,
+            use_safetensors=True,
+            attn_implementation="eager",
+            torch_dtype=torch.bfloat16,
+        ).eval().to("cuda")
+
+        try:
+            if getattr(tokenizer, "pad_token_id", None) is None and getattr(tokenizer, "eos_token_id", None) is not None:
+                tokenizer.pad_token = tokenizer.eos_token
+            if getattr(model.config, "pad_token_id", None) is None and getattr(tokenizer, "pad_token_id", None) is not None:
+                model.config.pad_token_id = tokenizer.pad_token_id
+        except Exception:
+            pass
+
+        self._model = model
+        self._tokenizer = tokenizer
+        print("✅ DeepSeek-OCR loaded and ready!")
+
+    def run(self, image_path, *, mode, prompt, grounding, find_term, schema, include_caption, options):
+        self._ensure_loaded()
+
+        prompt_text = build_prompt(
+            mode=mode,
+            user_prompt=prompt,
+            grounding=grounding,
+            find_term=find_term,
+            schema=schema,
+            include_caption=include_caption,
+        )
+
+        out_dir = tempfile.mkdtemp(prefix="dsocr_")
+        try:
+            res = self._model.infer(
+                self._tokenizer,
+                prompt=prompt_text,
+                image_file=image_path,
+                output_path=out_dir,
+                base_size=int(options.get("base_size", 1024)),
+                image_size=int(options.get("image_size", 640)),
+                crop_mode=bool(options.get("crop_mode", True)),
+                save_results=False,
+                test_compress=bool(options.get("test_compress", False)),
+                eval_mode=True,
+            )
+
+            if isinstance(res, str):
+                text = res.strip()
+            elif isinstance(res, dict) and "text" in res:
+                text = str(res["text"]).strip()
+            elif isinstance(res, (list, tuple)):
+                text = "\n".join(map(str, res)).strip()
+            else:
+                text = ""
+
+            if not text:
+                mmd = os.path.join(out_dir, "result.mmd")
+                if os.path.exists(mmd):
+                    with open(mmd, "r", encoding="utf-8") as fh:
+                        text = fh.read().strip()
+            return text
+        finally:
+            shutil.rmtree(out_dir, ignore_errors=True)
+
+
+class OllamaProvider(OCRProvider):
+    """A single vision model served by an external Ollama host."""
+
+    def __init__(self, tag: str, base_url: str, label: Optional[str] = None):
+        self.tag = tag
+        self.base_url = base_url.rstrip("/")
+        self.id = f"ollama:{tag}"
+        self.label = label or f"{tag} (Ollama)"
+        # Generic vision models don't emit DeepSeek grounding tokens.
+        self.capabilities = {"grounding": False, "advanced_settings": False}
+
+    def run(self, image_path, *, mode, prompt, grounding, find_term, schema, include_caption, options):
+        if httpx is None:
+            raise ProviderError("httpx is not installed; cannot reach Ollama.")
+
+        prompt_text = build_ollama_prompt(
+            mode=mode,
+            user_prompt=prompt,
+            find_term=find_term,
+            schema=schema,
+            include_caption=include_caption,
+        )
+
+        with open(image_path, "rb") as f:
+            img_b64 = base64.b64encode(f.read()).decode("utf-8")
+
+        payload = {
+            "model": self.tag,
+            "prompt": prompt_text,
+            "images": [img_b64],
+            "stream": False,
+        }
+
+        timeout = float(env_config("OLLAMA_TIMEOUT", default=300.0, cast=float))
+        try:
+            resp = httpx.post(f"{self.base_url}/api/generate", json=payload, timeout=timeout)
+            resp.raise_for_status()
+            data = resp.json()
+        except httpx.HTTPStatusError as e:
+            detail = ""
+            try:
+                detail = e.response.json().get("error", "")
+            except Exception:
+                detail = e.response.text[:200]
+            raise ProviderError(f"Ollama returned {e.response.status_code}: {detail}") from e
+        except httpx.HTTPError as e:
+            raise ProviderError(f"Could not reach Ollama at {self.base_url}: {e}") from e
+
+        return (data.get("response") or "").strip()
+
+
+# =============================================================================
+# Registry
+# =============================================================================
+class ModelRegistry:
+    def __init__(self, providers: List[OCRProvider], default_id: str):
+        self._providers: Dict[str, OCRProvider] = {p.id: p for p in providers}
+        # Fall back to the first registered provider if the configured default is gone.
+        self.default_id = default_id if default_id in self._providers else (
+            next(iter(self._providers), None)
+        )
+
+    def get(self, model_id: Optional[str]) -> OCRProvider:
+        chosen = model_id or self.default_id
+        provider = self._providers.get(chosen)
+        if provider is None:
+            raise ProviderError(f"Unknown model '{chosen}'.")
+        return provider
+
+    def list_models(self) -> List[Dict[str, Any]]:
+        out = []
+        for p in self._providers.values():
+            entry = p.info()
+            entry["default"] = (p.id == self.default_id)
+            out.append(entry)
+        return out
+
+
+def build_registry() -> ModelRegistry:
+    """Build the provider registry from environment variables.
+
+    Env:
+      ENABLE_DEEPSEEK_LOCAL  - register the local DeepSeek-OCR model (default: true)
+      OLLAMA_BASE_URL        - Ollama host (default: http://host.docker.internal:11434)
+      OLLAMA_MODELS          - comma-separated tags to surface (e.g. "glm-ocr,llama3.2-vision")
+      DEFAULT_OCR_MODEL      - id to select by default (default: deepseek-local)
+    """
+    providers: List[OCRProvider] = []
+
+    enable_deepseek = env_config("ENABLE_DEEPSEEK_LOCAL", default="true").strip().lower() in {"1", "true", "yes"}
+    if enable_deepseek:
+        providers.append(DeepSeekLocalProvider())
+
+    base_url = env_config("OLLAMA_BASE_URL", default="http://host.docker.internal:11434")
+    raw_tags = env_config("OLLAMA_MODELS", default="")
+    tags = [t.strip() for t in raw_tags.split(",") if t.strip()]
+    for tag in tags:
+        providers.append(OllamaProvider(tag=tag, base_url=base_url))
+
+    default_id = env_config("DEFAULT_OCR_MODEL", default="deepseek-local")
+    if not providers:
+        # Defensive: nothing configured. Register DeepSeek so the app still starts.
+        providers.append(DeepSeekLocalProvider())
+        default_id = "deepseek-local"
+
+    registry = ModelRegistry(providers, default_id)
+    print(f"🧠 OCR models registered: {[p.id for p in providers]} (default: {registry.default_id})")
+    return registry

backend/requirements.txt

@@ -15,3 +15,5 @@ PyMuPDF>=1.23.0
 img2pdf>=0.5.0
 python-docx>=1.1.0
 markdown>=3.5.0
+psycopg2-binary>=2.9.0
+httpx>=0.27.0

docker-compose.yml

@@ -1,4 +1,19 @@
 services:
+  postgres:
+    image: postgres:16-alpine
+    container_name: deepseek-ocr-postgres
+    environment:
+      POSTGRES_USER: ${POSTGRES_USER:-ocr_user}
+      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:-ocr_password}
+      POSTGRES_DB: ${POSTGRES_DB:-ocr_db}
+    volumes:
+      - postgres_data:/var/lib/postgresql/data
+    healthcheck:
+      test: ["CMD-SHELL", "pg_isready -U ${POSTGRES_USER:-ocr_user} -d ${POSTGRES_DB:-ocr_db}"]
+      interval: 5s
+      timeout: 5s
+      retries: 10
+
   backend:
     build: ./backend
     container_name: deepseek-ocr-backend
@@ -10,8 +25,23 @@ services:
       API_HOST: ${API_HOST:-0.0.0.0}
       API_PORT: ${API_PORT:-8000}
       MAX_UPLOAD_SIZE_MB: ${MAX_UPLOAD_SIZE_MB:-100}
+      DATABASE_URL: ${DATABASE_URL:-postgresql://ocr_user:ocr_password@postgres:5432/ocr_db}
+      OCR_IMAGES_DIR: ${OCR_IMAGES_DIR:-/data/ocr_images}
+      ENABLE_DEEPSEEK_LOCAL: ${ENABLE_DEEPSEEK_LOCAL:-true}
+      OLLAMA_BASE_URL: ${OLLAMA_BASE_URL:-http://host.docker.internal:11434}
+      OLLAMA_MODELS: ${OLLAMA_MODELS:-}
+      DEFAULT_OCR_MODEL: ${DEFAULT_OCR_MODEL:-deepseek-local}
+      OLLAMA_TIMEOUT: ${OLLAMA_TIMEOUT:-300}
+    # Lets the container reach an Ollama server running on the Docker host
+    # (works out of the box on Docker Desktop; required for Linux engines).
+    extra_hosts:
+      - "host.docker.internal:host-gateway"
     volumes:
       - ./models:/models
+      - ./ocr_images:/data/ocr_images
+    depends_on:
+      postgres:
+        condition: service_healthy
     deploy:
       resources:
         reservations:
@@ -22,8 +52,6 @@ services:
     shm_size: "4g"
     ports:
       - "${API_PORT:-8000}:${API_PORT:-8000}"
-    networks:
-      - ocr-network
 
   frontend:
     build: ./frontend
@@ -32,9 +60,10 @@ services:
       - "${FRONTEND_PORT:-3000}:80"
     depends_on:
       - backend
-    networks:
+
-      - ocr-network
+volumes:
+  postgres_data:
 
 networks:
-  ocr-network:
+  default:
-    driver: bridge
+    name: rw-research

frontend/src/App.jsx

@@ -1,18 +1,35 @@
-import { useState, useCallback } from 'react'
+import { useState, useCallback, useEffect } from 'react'
+import { useSuggestions } from './hooks/useSuggestions'
+import { useModels } from './hooks/useModels'
 import { motion, AnimatePresence } from 'framer-motion'
-import { Sparkles, Zap, Loader2, Settings, Image as ImageIcon, FileText } from 'lucide-react'
+import {
+  Sparkles, Zap, Loader2, Settings, Image as ImageIcon, FileText,
+  Layers, ChevronLeft, CheckCircle2, Database,
+} from 'lucide-react'
 import ImageUpload from './components/ImageUpload'
 import ModeSelector from './components/ModeSelector'
+import ModelSelector from './components/ModelSelector'
 import ResultPanel from './components/ResultPanel'
 import AdvancedSettings from './components/AdvancedSettings'
 import PDFProcessor from './components/PDFProcessor'
+import MetadataForm from './components/MetadataForm'
+import JobsPanel from './components/JobsPanel'
 import axios from 'axios'
 
 const API_BASE = import.meta.env.VITE_API_URL || '/api'
 
+const INPUT_CLASS =
+  'w-full bg-white/5 border border-white/10 rounded-lg px-3 py-2 text-sm text-gray-200 ' +
+  'placeholder-gray-600 focus:outline-none focus:border-purple-500/50 transition-colors'
+
 function App() {
+  const [view, setView] = useState('new_job')
+
+  // OCR state
+  const { models, loading: modelsLoading } = useModels()
+  const [model, setModel] = useState(null)
   const [mode, setMode] = useState('plain_ocr')
-  const [fileType, setFileType] = useState('image') // 'image' or 'pdf'
+  const [fileType, setFileType] = useState('image')
   const [image, setImage] = useState(null)
   const [imagePreview, setImagePreview] = useState(null)
   const [result, setResult] = useState(null)
@@ -21,22 +38,39 @@ function App() {
   const [showAdvanced, setShowAdvanced] = useState(false)
   const [includeCaption, setIncludeCaption] = useState(false)
 
-  // Form state
   const [prompt, setPrompt] = useState('')
   const [findTerm, setFindTerm] = useState('')
   const [advancedSettings, setAdvancedSettings] = useState({
-    base_size: 1024,
+    base_size: 1024, image_size: 640, crop_mode: true, test_compress: false,
-    image_size: 640,
-    crop_mode: true,
-    test_compress: false
   })
 
-  const handleFileTypeChange = useCallback((newType) => {
+  const suggestions = useSuggestions()
-    // Clear current file when switching types
+
-    setImage(null)
+  const [metadata, setMetadata] = useState({ author: '', book: '', chapter: '', page: '' })
-    if (imagePreview) {
+  // Results accumulated per mode: { plain_ocr: 'text', describe: 'text', freeform: 'text' }
-      URL.revokeObjectURL(imagePreview)
+  const [modeResults, setModeResults] = useState({})
+  const [editedResults, setEditedResults] = useState({})
+  const [activeResultMode, setActiveResultMode] = useState(null)
+  const [commitLoading, setCommitLoading] = useState(false)
+  const [commitResult, setCommitResult] = useState(null)
+
+  // Modes that produce editable text output and can be committed to the DB
+  const COMMITTABLE_MODES = new Set(['plain_ocr', 'describe'])
+  const MODE_LABELS = { plain_ocr: 'OCR Text', describe: 'Description' }
+
+  // Pick the default model once the list loads
+  useEffect(() => {
+    if (!model && models.length > 0) {
+      setModel((models.find(m => m.default) || models[0]).id)
     }
+  }, [models, model])
+
+  // Show the full-screen result view once at least one committable mode has a result
+  const showResultView = view === 'new_job' && Object.keys(modeResults).length > 0
+
+  const handleFileTypeChange = useCallback((newType) => {
+    setImage(null)
+    if (imagePreview) URL.revokeObjectURL(imagePreview)
     setImagePreview(null)
     setError(null)
     setResult(null)
@@ -45,42 +79,38 @@ function App() {
 
   const handleImageSelect = useCallback((file) => {
     if (file === null) {
-      // Clear everything when removing image
       setImage(null)
-      if (imagePreview && fileType === 'image') {
+      if (imagePreview && fileType === 'image') URL.revokeObjectURL(imagePreview)
-        URL.revokeObjectURL(imagePreview)
-      }
       setImagePreview(null)
       setError(null)
       setResult(null)
+      setModeResults({})
+      setEditedResults({})
+      setActiveResultMode(null)
+      setCommitResult(null)
     } else {
       setImage(file)
-      // Only create preview URL for images, not PDFs
+      setImagePreview(fileType === 'image' ? URL.createObjectURL(file) : file)
-      if (fileType === 'image') {
-        setImagePreview(URL.createObjectURL(file))
-      } else {
-        setImagePreview(file) // Just store the file for PDFs
-      }
       setError(null)
       setResult(null)
+      setModeResults({})
+      setEditedResults({})
+      setActiveResultMode(null)
+      setCommitResult(null)
     }
   }, [imagePreview, fileType])
 
   const handleSubmit = async () => {
-    if (!image) {
+    if (!image) { setError('Please upload an image first'); return }
-      setError('Please upload an image first')
-      return
-    }
-
     setLoading(true)
     setError(null)
-
+    setCommitResult(null)
     try {
       const formData = new FormData()
       formData.append('image', image)
+      if (model) formData.append('model', model)
       formData.append('mode', mode)
       formData.append('prompt', prompt)
-      // Enable grounding only for find mode
       formData.append('grounding', mode === 'find_ref')
       formData.append('include_caption', includeCaption)
       formData.append('find_term', findTerm)
@@ -91,12 +121,16 @@ function App() {
       formData.append('test_compress', advancedSettings.test_compress)
 
       const response = await axios.post(`${API_BASE}/ocr`, formData, {
-        headers: {
+        headers: { 'Content-Type': 'multipart/form-data' },
-          'Content-Type': 'multipart/form-data',
-        },
       })
-
       setResult(response.data)
+      if (COMMITTABLE_MODES.has(mode)) {
+        const text = response.data.text || ''
+        setModeResults(prev => ({ ...prev, [mode]: text }))
+        setEditedResults(prev => ({ ...prev, [mode]: text }))
+        setActiveResultMode(mode)
+      }
+      setCommitResult(null)
     } catch (err) {
       setError(err.response?.data?.detail || err.message || 'An error occurred')
     } finally {
@@ -104,31 +138,61 @@ function App() {
     }
   }
 
-  const handleCopy = useCallback(() => {
+  const handleNewAnalysis = () => {
-    if (result?.text) {
+    setResult(null)
-      navigator.clipboard.writeText(result.text)
+    setModeResults({})
+    setEditedResults({})
+    setActiveResultMode(null)
+    setCommitResult(null)
+  }
+
+  const handleCommitJob = useCallback(async () => {
+    if (!image) return
+    setCommitLoading(true)
+    setCommitResult(null)
+    try {
+      const formData = new FormData()
+      formData.append('image', image)
+      formData.append('author', metadata.author)
+      formData.append('book', metadata.book)
+      formData.append('chapter', metadata.chapter)
+      formData.append('page', metadata.page)
+      formData.append('ocr_text', editedResults.plain_ocr || '')
+      formData.append('describe_text', editedResults.describe || '')
+      formData.append('freeform_text', editedResults.freeform || '')
+      formData.append('mode', mode)
+      if (model) formData.append('ocr_model', model)
+
+      const response = await axios.post(`${API_BASE}/jobs`, formData, {
+        headers: { 'Content-Type': 'multipart/form-data' },
+      })
+      setCommitResult({ success: true, job: response.data })
+    } catch (err) {
+      setCommitResult({ success: false, error: err.response?.data?.detail || err.message })
+    } finally {
+      setCommitLoading(false)
     }
-  }, [result])
+  }, [image, editedResults, metadata, mode, model])
+
+  const handleCopy = useCallback(() => {
+    const text = (activeResultMode && editedResults[activeResultMode]) || result?.text
+    if (text) navigator.clipboard.writeText(text)
+  }, [activeResultMode, editedResults, result])
 
   const handleDownload = useCallback(() => {
-    if (!result?.text) return
+    const text = (activeResultMode && editedResults[activeResultMode]) || result?.text
-    
+    if (!text) return
-    const extensions = {
+    const ext = { plain_ocr: 'txt', describe: 'txt', find_ref: 'txt', freeform: 'txt' }[mode] || 'txt'
-      plain_ocr: 'txt',
+    const blob = new Blob([text], { type: 'text/plain' })
-      describe: 'txt',
-      find_ref: 'txt',
-      freeform: 'txt',
-    }
-    
-    const ext = extensions[mode] || 'txt'
-    const blob = new Blob([result.text], { type: 'text/plain' })
     const url = URL.createObjectURL(blob)
     const a = document.createElement('a')
     a.href = url
     a.download = `deepseek-ocr-result.${ext}`
     a.click()
     URL.revokeObjectURL(url)
-  }, [result, mode])
+  }, [activeResultMode, editedResults, result, mode])
+
+  const metaField = (key) => (e) => setMetadata(m => ({ ...m, [key]: e.target.value }))
 
   return (
     <div className="min-h-screen relative overflow-hidden">
@@ -138,27 +202,13 @@ function App() {
         <div className="absolute inset-0 bg-[url('data:image/svg+xml;base64,PHN2ZyB3aWR0aD0iNjAiIGhlaWdodD0iNjAiIHZpZXdCb3g9IjAgMCA2MCA2MCIgeG1sbnM9Imh0dHA6Ly93d3cudzMub3JnLzIwMDAvc3ZnIj48ZyBmaWxsPSJub25lIiBmaWxsLXJ1bGU9ImV2ZW5vZGQiPjxwYXRoIGQ9Ik0zNiAxOGMzLjMxIDAgNiAyLjY5IDYgNnMtMi42OSA2LTYgNi02LTIuNjktNi02IDIuNjktNiA2LTZ6TTI0IDZjMy4zMSAwIDYgMi42OSA2IDZzLTIuNjkgNi02IDYtNi0yLjY5LTYtNiAyLjY5LTYgNi02ek00OCAzNmMzLjMxIDAgNiAyLjY5IDYgNnMtMi42OSA2LTYgNi02LTIuNjktNi02IDIuNjktNiA2LTZ6IiBzdHJva2U9InJnYmEoMTQ3LCA1MSwgMjM0LCAwLjEpIiBzdHJva2Utd2lkdGg9IjIiLz48L2c+PC9zdmc+')] opacity-30" />
         <motion.div
           className="absolute top-20 left-20 w-96 h-96 bg-purple-500/10 rounded-full blur-3xl"
-          animate={{
+          animate={{ scale: [1, 1.2, 1], opacity: [0.3, 0.5, 0.3] }}
-            scale: [1, 1.2, 1],
+          transition={{ duration: 8, repeat: Infinity, ease: 'easeInOut' }}
-            opacity: [0.3, 0.5, 0.3],
-          }}
-          transition={{
-            duration: 8,
-            repeat: Infinity,
-            ease: "easeInOut"
-          }}
         />
         <motion.div
           className="absolute bottom-20 right-20 w-96 h-96 bg-cyan-500/10 rounded-full blur-3xl"
-          animate={{
+          animate={{ scale: [1.2, 1, 1.2], opacity: [0.5, 0.3, 0.5] }}
-            scale: [1.2, 1, 1.2],
+          transition={{ duration: 8, repeat: Infinity, ease: 'easeInOut' }}
-            opacity: [0.5, 0.3, 0.5],
-          }}
-          transition={{
-            duration: 8,
-            repeat: Infinity,
-            ease: "easeInOut"
-          }}
         />
       </div>
 
@@ -166,11 +216,7 @@ function App() {
       <header className="sticky top-0 z-50 glass border-b border-white/10">
         <div className="max-w-7xl mx-auto px-6 py-4">
           <div className="flex items-center justify-between">
-            <motion.div 
+            <motion.div className="flex items-center gap-3" initial={{ opacity: 0, x: -20 }} animate={{ opacity: 1, x: 0 }}>
-              className="flex items-center gap-3"
-              initial={{ opacity: 0, x: -20 }}
-              animate={{ opacity: 1, x: 0 }}
-            >
               <div className="relative">
                 <div className="absolute inset-0 bg-gradient-to-r from-purple-500 to-cyan-500 rounded-xl blur-lg opacity-75" />
                 <div className="relative bg-gradient-to-br from-purple-600 to-cyan-500 p-2 rounded-xl">
@@ -182,173 +228,348 @@ function App() {
                 <p className="text-xs text-gray-400">Next-Gen Vision AI</p>
               </div>
             </motion.div>
+
+            <nav className="flex gap-2">
+              {showResultView && (
+                <motion.button
+                  onClick={handleNewAnalysis}
+                  className="flex items-center gap-2 px-4 py-2 rounded-xl text-sm font-medium glass text-gray-400 hover:bg-white/5 transition-all"
+                  whileHover={{ scale: 1.02 }} whileTap={{ scale: 0.98 }}
+                >
+                  <ChevronLeft className="w-4 h-4" />
+                  New Analysis
+                </motion.button>
+              )}
+              <motion.button
+                onClick={() => setView('new_job')}
+                className={`flex items-center gap-2 px-4 py-2 rounded-xl text-sm font-medium transition-all ${view === 'new_job' ? 'bg-gradient-to-r from-purple-600 to-cyan-600 text-white' : 'glass text-gray-400 hover:bg-white/5'}`}
+                whileHover={{ scale: 1.02 }} whileTap={{ scale: 0.98 }}
+              >
+                <Zap className="w-4 h-4" />
+                New Job
+              </motion.button>
+              <motion.button
+                onClick={() => setView('jobs')}
+                className={`flex items-center gap-2 px-4 py-2 rounded-xl text-sm font-medium transition-all ${view === 'jobs' ? 'bg-gradient-to-r from-purple-600 to-cyan-600 text-white' : 'glass text-gray-400 hover:bg-white/5'}`}
+                whileHover={{ scale: 1.02 }} whileTap={{ scale: 0.98 }}
+              >
+                <Layers className="w-4 h-4" />
+                Browse Jobs
+              </motion.button>
+            </nav>
           </div>
         </div>
       </header>
 
       {/* Main Content */}
-      <main className="max-w-7xl mx-auto px-6 py-8">
+      <main className="max-w-7xl mx-auto px-6 py-6">
-        <div className="grid lg:grid-cols-2 gap-6">
+        <AnimatePresence>
-          {/* Left Panel - Upload & Controls */}
-          <motion.div
-            initial={{ opacity: 0, y: 20 }}
-            animate={{ opacity: 1, y: 0 }}
-            transition={{ delay: 0.1 }}
-            className="space-y-6"
-          >
-            {/* File Type Toggle */}
-            <div className="glass p-4 rounded-2xl">
-              <div className="grid grid-cols-2 gap-2">
-                <motion.button
-                  onClick={() => handleFileTypeChange('image')}
-                  className={`p-3 rounded-xl text-sm font-medium transition-all flex items-center justify-center gap-2 ${
-                    fileType === 'image'
-                      ? 'bg-gradient-to-r from-purple-600 to-cyan-600 text-white'
-                      : 'glass text-gray-400 hover:bg-white/5'
-                  }`}
-                  whileHover={{ scale: 1.02 }}
-                  whileTap={{ scale: 0.98 }}
-                >
-                  <ImageIcon className="w-4 h-4" />
-                  Image OCR
-                </motion.button>
-                <motion.button
-                  onClick={() => handleFileTypeChange('pdf')}
-                  className={`p-3 rounded-xl text-sm font-medium transition-all flex items-center justify-center gap-2 ${
-                    fileType === 'pdf'
-                      ? 'bg-gradient-to-r from-purple-600 to-cyan-600 text-white'
-                      : 'glass text-gray-400 hover:bg-white/5'
-                  }`}
-                  whileHover={{ scale: 1.02 }}
-                  whileTap={{ scale: 0.98 }}
-                >
-                  <FileText className="w-4 h-4" />
-                  PDF Processing
-                </motion.button>
-              </div>
-            </div>
 
-            {/* Mode Selector with integrated inputs */}
+          {/* ── Full-screen OCR result view ── */}
-            <ModeSelector
+          {showResultView ? (
-              mode={mode}
+            <motion.div
-              onModeChange={setMode}
+              key="ocr_result"
-              prompt={prompt}
+              initial={{ opacity: 0, y: 20 }}
-              onPromptChange={setPrompt}
+              animate={{ opacity: 1, y: 0 }}
-              findTerm={findTerm}
+              exit={{ opacity: 0, y: -20 }}
-              onFindTermChange={setFindTerm}
+              className="flex flex-col gap-4"
-            />
-
-            {/* Image/PDF Upload */}
-            <ImageUpload
-              onImageSelect={handleImageSelect}
-              preview={imagePreview}
-              fileType={fileType}
-            />
-
-            {/* Advanced Settings Toggle */}
-            <motion.button
-              onClick={() => setShowAdvanced(!showAdvanced)}
-              className="w-full glass px-4 py-3 rounded-2xl flex items-center justify-between hover:bg-white/5 transition-colors"
-              whileHover={{ scale: 1.01 }}
-              whileTap={{ scale: 0.99 }}
             >
-              <div className="flex items-center gap-2">
+              {/* Run additional modes */}
-                <Settings className="w-4 h-4 text-purple-400" />
+              <div className="glass p-4 rounded-2xl flex-shrink-0">
-                <span className="text-sm font-medium text-gray-300">Advanced Settings</span>
+                <div className="mb-3">
-              </div>
+                  <ModelSelector
-              <motion.div
+                    models={models} value={model} onChange={setModel} loading={modelsLoading}
-                animate={{ rotate: showAdvanced ? 180 : 0 }}
+                  />
-                transition={{ duration: 0.3 }}
+                </div>
-              >
+                <ModeSelector mode={mode} onModeChange={setMode} />
-                <svg className="w-4 h-4 text-gray-400" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+                <div className="flex items-center gap-3 mt-3">
-                  <path strokeLinecap="round" strokeLinejoin="round" strokeWidth={2} d="M19 9l-7 7-7-7" />
+                  <motion.button
-                </svg>
+                    onClick={handleSubmit}
-              </motion.div>
+                    disabled={loading}
-            </motion.button>
+                    className={`flex items-center gap-2 px-5 py-2 rounded-xl font-medium text-sm transition-all ${loading ? 'opacity-50 cursor-not-allowed bg-white/5' : 'bg-gradient-to-r from-purple-600 to-cyan-600'}`}
-
+                    whileHover={!loading ? { scale: 1.02 } : {}}
-            {/* Advanced Settings Panel */}
+                    whileTap={!loading ? { scale: 0.98 } : {}}
-            <AnimatePresence>
-              {showAdvanced && (
-                <AdvancedSettings
-                  settings={advancedSettings}
-                  onSettingsChange={setAdvancedSettings}
-                  includeCaption={includeCaption}
-                  onIncludeCaptionChange={setIncludeCaption}
-                />
-              )}
-            </AnimatePresence>
-
-            {/* Action Button / PDF Processor */}
-            {fileType === 'pdf' ? (
-              <PDFProcessor
-                pdfFile={image}
-                mode={mode}
-                prompt={prompt}
-                advancedSettings={advancedSettings}
-                includeCaption={includeCaption}
-              />
-            ) : (
-              <>
-                <motion.button
-                  onClick={handleSubmit}
-                  disabled={!image || loading}
-                  className={`w-full relative overflow-hidden rounded-2xl p-[2px] ${
-                    !image || loading ? 'opacity-50 cursor-not-allowed' : ''
-                  }`}
-                  whileHover={!loading && image ? { scale: 1.02 } : {}}
-                  whileTap={!loading && image ? { scale: 0.98 } : {}}
-                >
-                  <div className="absolute inset-0 bg-gradient-to-r from-purple-600 via-pink-600 to-cyan-600 animate-gradient" />
-                  <div className="relative bg-dark-100 px-8 py-4 rounded-2xl flex items-center justify-center gap-3">
-                    {loading ? (
-                      <>
-                        <Loader2 className="w-5 h-5 animate-spin" />
-                        <span className="font-semibold">Processing Magic...</span>
-                      </>
-                    ) : (
-                      <>
-                        <Zap className="w-5 h-5" />
-                        <span className="font-semibold">Analyze Image</span>
-                      </>
-                    )}
-                  </div>
-                </motion.button>
-
-                {error && (
-                  <motion.div
-                    initial={{ opacity: 0, y: -10 }}
-                    animate={{ opacity: 1, y: 0 }}
-                    className="glass p-4 rounded-2xl border-red-500/50 bg-red-500/10"
                   >
-                    <p className="text-sm text-red-400">{error}</p>
+                    {loading
-                  </motion.div>
+                      ? <><Loader2 className="w-4 h-4 animate-spin" /> Processing...</>
-                )}
+                      : <><Zap className="w-4 h-4" /> Analyze</>}
-              </>
+                  </motion.button>
-            )}
+                  {error && <p className="text-sm text-red-400">{error}</p>}
-          </motion.div>
+                </div>
+              </div>
 
-          {/* Right Panel - Results */}
+              {/* Image + Text */}
-          <motion.div
+              <div className="grid gap-6" style={{ gridTemplateColumns: '1fr 1fr', height: '130vh' }}>
-            initial={{ opacity: 0, y: 20 }}
+                {imagePreview && typeof imagePreview === 'string' ? (
-            animate={{ opacity: 1, y: 0 }}
+                  <div className="glass rounded-2xl overflow-hidden flex items-center justify-center bg-black/20 h-full">
-            transition={{ delay: 0.2 }}
+                    <img
-          >
+                      src={imagePreview}
-            <ResultPanel 
+                      alt="Source"
-              result={result}
+                      className="w-full h-full object-contain"
-              loading={loading}
+                    />
-              imagePreview={imagePreview}
+                  </div>
-              onCopy={handleCopy}
+                ) : (
-              onDownload={handleDownload}
+                  <div className="glass rounded-2xl flex items-center justify-center h-full">
-            />
+                    <p className="text-gray-500 text-sm">No preview</p>
-          </motion.div>
+                  </div>
-        </div>
+                )}
+                <div className="glass rounded-2xl p-4 flex flex-col h-full">
+                  {/* Mode tabs — only shown when multiple modes have results */}
+                  {Object.keys(modeResults).length > 1 && (
+                    <div className="flex gap-1 mb-3 flex-shrink-0">
+                      {Object.keys(modeResults).map(m => (
+                        <button
+                          key={m}
+                          onClick={() => setActiveResultMode(m)}
+                          className={`px-3 py-1 rounded-lg text-xs font-medium transition-colors ${
+                            activeResultMode === m
+                              ? 'bg-purple-600 text-white'
+                              : 'bg-white/5 text-gray-400 hover:bg-white/10'
+                          }`}
+                        >
+                          {MODE_LABELS[m] || m}
+                        </button>
+                      ))}
+                    </div>
+                  )}
+                  <p className="text-xs text-gray-400 mb-2 flex-shrink-0">
+                    {MODE_LABELS[activeResultMode] || 'Result'}
+                    <span className="text-purple-400 ml-1">(edit before committing)</span>
+                  </p>
+                  {loading && COMMITTABLE_MODES.has(mode) ? (
+                    <div className="flex-1 flex items-center justify-center">
+                      <Loader2 className="w-8 h-8 animate-spin text-purple-400" />
+                    </div>
+                  ) : (
+                    <textarea
+                      value={activeResultMode ? (editedResults[activeResultMode] ?? '') : ''}
+                      onChange={e => setEditedResults(prev => ({ ...prev, [activeResultMode]: e.target.value }))}
+                      className="flex-1 w-full bg-transparent text-sm text-gray-200 font-mono resize-none focus:outline-none min-h-0"
+                      placeholder="Run a mode to see results here..."
+                    />
+                  )}
+                </div>
+              </div>
+
+              {/* Metadata row */}
+              <div className="glass p-4 rounded-2xl flex-shrink-0">
+                <datalist id="rv-authors">
+                  {suggestions.authors.map(a => <option key={a} value={a} />)}
+                </datalist>
+                <datalist id="rv-books">
+                  {(suggestions.books || []).map(b => <option key={b} value={b} />)}
+                </datalist>
+                <datalist id="rv-chapters">
+                  {suggestions.chapters.map(c => <option key={c} value={c} />)}
+                </datalist>
+                <div className="grid grid-cols-4 gap-4">
+                  {[
+                    { key: 'author',  label: 'Author',  placeholder: 'Author name',  list: 'rv-authors' },
+                    { key: 'book',    label: 'Book',    placeholder: 'Book title',    list: 'rv-books' },
+                    { key: 'chapter', label: 'Chapter', placeholder: 'Chapter',       list: 'rv-chapters' },
+                    { key: 'page',    label: 'Page',    placeholder: 'Page number',   list: undefined },
+                  ].map(({ key, label, placeholder, list }) => (
+                    <div key={key}>
+                      <label className="text-xs text-gray-400 mb-1 block">{label}</label>
+                      <input
+                        type="text"
+                        list={list}
+                        value={metadata[key]}
+                        onChange={metaField(key)}
+                        placeholder={placeholder}
+                        className={INPUT_CLASS}
+                      />
+                    </div>
+                  ))}
+                </div>
+              </div>
+
+              {/* Commit row */}
+              <div className="flex items-center gap-4 flex-shrink-0">
+                <AnimatePresence>
+                  {commitResult?.success && (
+                    <motion.div
+                      initial={{ opacity: 0, x: -10 }} animate={{ opacity: 1, x: 0 }} exit={{ opacity: 0 }}
+                      className="flex-1 glass p-3 rounded-xl bg-green-500/10 border border-green-500/20"
+                    >
+                      <p className="text-xs text-green-400">
+                        Job saved &mdash; ID: <span className="font-mono">{commitResult.job?.id}</span>
+                      </p>
+                    </motion.div>
+                  )}
+                  {commitResult && !commitResult.success && (
+                    <motion.div
+                      initial={{ opacity: 0, x: -10 }} animate={{ opacity: 1, x: 0 }} exit={{ opacity: 0 }}
+                      className="flex-1 glass p-3 rounded-xl bg-red-500/10 border border-red-500/20"
+                    >
+                      <p className="text-xs text-red-400">{commitResult.error}</p>
+                    </motion.div>
+                  )}
+                </AnimatePresence>
+                <motion.button
+                  onClick={handleCommitJob}
+                  disabled={commitLoading || commitResult?.success}
+                  className={`flex items-center gap-2 px-6 py-3 rounded-xl font-medium text-sm transition-all flex-shrink-0 ${
+                    commitLoading || commitResult?.success
+                      ? 'opacity-50 cursor-not-allowed bg-white/5'
+                      : 'bg-gradient-to-r from-blue-600 to-indigo-600 hover:from-blue-500 hover:to-indigo-500'
+                  }`}
+                  whileHover={!commitLoading && !commitResult?.success ? { scale: 1.02 } : {}}
+                  whileTap={!commitLoading && !commitResult?.success ? { scale: 0.98 } : {}}
+                >
+                  {commitLoading ? (
+                    <><Loader2 className="w-4 h-4 animate-spin" /> Committing...</>
+                  ) : commitResult?.success ? (
+                    <><CheckCircle2 className="w-4 h-4" /> Committed</>
+                  ) : (
+                    <><Database className="w-4 h-4" /> Commit Job</>
+                  )}
+                </motion.button>
+              </div>
+            </motion.div>
+
+          ) : view === 'jobs' ? (
+            <motion.div
+              key="jobs"
+              initial={{ opacity: 0, y: 20 }}
+              animate={{ opacity: 1, y: 0 }}
+              exit={{ opacity: 0, y: -20 }}
+            >
+              <JobsPanel />
+            </motion.div>
+
+          ) : (
+            /* ── Upload / Controls layout ── */
+            <motion.div
+              key="new_job"
+              initial={{ opacity: 0, y: 20 }}
+              animate={{ opacity: 1, y: 0 }}
+              exit={{ opacity: 0, y: -20 }}
+            >
+              <div className="grid lg:grid-cols-2 gap-6">
+                {/* Left Panel */}
+                <motion.div
+                  initial={{ opacity: 0, y: 20 }}
+                  animate={{ opacity: 1, y: 0 }}
+                  transition={{ delay: 0.1 }}
+                  className="space-y-6"
+                >
+                  {/* File Type Toggle */}
+                  <div className="glass p-4 rounded-2xl">
+                    <div className="grid grid-cols-2 gap-2">
+                      <motion.button
+                        onClick={() => handleFileTypeChange('image')}
+                        className={`p-3 rounded-xl text-sm font-medium transition-all flex items-center justify-center gap-2 ${fileType === 'image' ? 'bg-gradient-to-r from-purple-600 to-cyan-600 text-white' : 'glass text-gray-400 hover:bg-white/5'}`}
+                        whileHover={{ scale: 1.02 }} whileTap={{ scale: 0.98 }}
+                      >
+                        <ImageIcon className="w-4 h-4" /> Image OCR
+                      </motion.button>
+                      <motion.button
+                        onClick={() => handleFileTypeChange('pdf')}
+                        className={`p-3 rounded-xl text-sm font-medium transition-all flex items-center justify-center gap-2 ${fileType === 'pdf' ? 'bg-gradient-to-r from-purple-600 to-cyan-600 text-white' : 'glass text-gray-400 hover:bg-white/5'}`}
+                        whileHover={{ scale: 1.02 }} whileTap={{ scale: 0.98 }}
+                      >
+                        <FileText className="w-4 h-4" /> PDF Processing
+                      </motion.button>
+                    </div>
+                  </div>
+
+                  <MetadataForm metadata={metadata} onChange={setMetadata} suggestions={suggestions} />
+
+                  <ModelSelector
+                    models={models} value={model} onChange={setModel} loading={modelsLoading}
+                  />
+
+                  <ModeSelector mode={mode} onModeChange={setMode} />
+
+                  <ImageUpload onImageSelect={handleImageSelect} preview={imagePreview} fileType={fileType} />
+
+                  <motion.button
+                    onClick={() => setShowAdvanced(!showAdvanced)}
+                    className="w-full glass px-4 py-3 rounded-2xl flex items-center justify-between hover:bg-white/5 transition-colors"
+                    whileHover={{ scale: 1.01 }} whileTap={{ scale: 0.99 }}
+                  >
+                    <div className="flex items-center gap-2">
+                      <Settings className="w-4 h-4 text-purple-400" />
+                      <span className="text-sm font-medium text-gray-300">Advanced Settings</span>
+                    </div>
+                    <motion.div animate={{ rotate: showAdvanced ? 180 : 0 }} transition={{ duration: 0.3 }}>
+                      <svg className="w-4 h-4 text-gray-400" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+                        <path strokeLinecap="round" strokeLinejoin="round" strokeWidth={2} d="M19 9l-7 7-7-7" />
+                      </svg>
+                    </motion.div>
+                  </motion.button>
+
+                  <AnimatePresence>
+                    {showAdvanced && (
+                      <AdvancedSettings
+                        settings={advancedSettings} onSettingsChange={setAdvancedSettings}
+                        includeCaption={includeCaption} onIncludeCaptionChange={setIncludeCaption}
+                      />
+                    )}
+                  </AnimatePresence>
+
+                  {fileType === 'pdf' ? (
+                    <PDFProcessor
+                      pdfFile={image} mode={mode} prompt={prompt} model={model}
+                      advancedSettings={advancedSettings} includeCaption={includeCaption}
+                    />
+                  ) : (
+                    <>
+                      <motion.button
+                        onClick={handleSubmit}
+                        disabled={!image || loading}
+                        className={`w-full relative overflow-hidden rounded-2xl p-[2px] ${!image || loading ? 'opacity-50 cursor-not-allowed' : ''}`}
+                        whileHover={!loading && image ? { scale: 1.02 } : {}}
+                        whileTap={!loading && image ? { scale: 0.98 } : {}}
+                      >
+                        <div className="absolute inset-0 bg-gradient-to-r from-purple-600 via-pink-600 to-cyan-600 animate-gradient" />
+                        <div className="relative bg-dark-100 px-8 py-4 rounded-2xl flex items-center justify-center gap-3">
+                          {loading ? (
+                            <><Loader2 className="w-5 h-5 animate-spin" /><span className="font-semibold">Processing Magic...</span></>
+                          ) : (
+                            <><Zap className="w-5 h-5" /><span className="font-semibold">Analyze Image</span></>
+                          )}
+                        </div>
+                      </motion.button>
+
+                      {error && (
+                        <motion.div
+                          initial={{ opacity: 0, y: -10 }} animate={{ opacity: 1, y: 0 }}
+                          className="glass p-4 rounded-2xl border-red-500/50 bg-red-500/10"
+                        >
+                          <p className="text-sm text-red-400">{error}</p>
+                        </motion.div>
+                      )}
+                    </>
+                  )}
+                </motion.div>
+
+                {/* Right Panel - Results (non-plain_ocr modes or loading) */}
+                <motion.div
+                  initial={{ opacity: 0, y: 20 }}
+                  animate={{ opacity: 1, y: 0 }}
+                  transition={{ delay: 0.2 }}
+                >
+                  <ResultPanel
+                    result={result}
+                    loading={loading}
+                    imagePreview={imagePreview}
+                    onCopy={handleCopy}
+                    onDownload={handleDownload}
+                  />
+                </motion.div>
+              </div>
+            </motion.div>
+          )}
+        </AnimatePresence>
       </main>
 
       {/* Footer */}
       <footer className="mt-20 border-t border-white/10 glass">
         <div className="max-w-7xl mx-auto px-6 py-8 text-center space-y-2">
           <p className="text-sm text-gray-400">
-            Powered by <span className="gradient-text font-semibold">DeepSeek-OCR</span> • 
+            Powered by <span className="gradient-text font-semibold">DeepSeek-OCR</span> &bull;
             Built with <span className="text-pink-400">♥</span> using React + FastAPI
           </p>
           <p className="text-xs text-gray-500">

frontend/src/components/JobsPanel.jsx

@@ -0,0 +1,665 @@
+import { useState, useEffect, useCallback } from 'react'
+import { useSuggestions } from '../hooks/useSuggestions'
+import { useModels } from '../hooks/useModels'
+import { motion, AnimatePresence } from 'framer-motion'
+import {
+  Search, ChevronLeft, ChevronRight, CheckCircle2, Clock,
+  FileText, Loader2, Save, RefreshCw, Trash2, Sparkles,
+} from 'lucide-react'
+import axios from 'axios'
+
+const API_BASE = import.meta.env.VITE_API_URL || '/api'
+
+const INPUT_CLASS =
+  'w-full bg-white/5 border border-white/10 rounded-lg px-3 py-2 text-sm text-gray-200 ' +
+  'placeholder-gray-600 focus:outline-none focus:border-purple-500/50 transition-colors'
+
+const STATUS_COLORS = {
+  unreviewed: 'text-amber-400 bg-amber-400/10 border-amber-400/30',
+  reviewed:   'text-green-400 bg-green-400/10 border-green-400/30',
+}
+
+function StatusBadge({ status }) {
+  const Icon = status === 'reviewed' ? CheckCircle2 : Clock
+  return (
+    <span className={`inline-flex items-center gap-1 px-2 py-0.5 rounded-full text-xs border ${STATUS_COLORS[status] || 'text-gray-400'}`}>
+      <Icon className="w-3 h-3" />
+      {status}
+    </span>
+  )
+}
+
+// ─────────────────────────────────────────────────────────────
+// Full-screen Job Detail
+// ─────────────────────────────────────────────────────────────
+function JobDetail({ jobId, onClose, onReviewed, onDeleted, suggestions = {} }) {
+  const { models } = useModels()
+  const [job, setJob] = useState(null)
+  const [loading, setLoading] = useState(true)
+  const [error, setError] = useState(null)
+
+  const [describeModel, setDescribeModel] = useState('')
+  const [generatingDescribe, setGeneratingDescribe] = useState(false)
+
+  const [editedText, setEditedText]         = useState('')
+  const [editDescribeText, setEditDescribeText] = useState('')
+  const [editFreeformText, setEditFreeformText] = useState('')
+  const [activeTab, setActiveTab]           = useState('ocr')
+  const [editAuthor, setEditAuthor]         = useState('')
+  const [editBook, setEditBook]             = useState('')
+  const [editChapter, setEditChapter]       = useState('')
+  const [editPage, setEditPage]             = useState('')
+  const [reviewerName, setReviewerName]     = useState('')
+
+  const [submitting, setSubmitting] = useState(false)
+  const [saveResult, setSaveResult] = useState(null)
+  const [confirmDelete, setConfirmDelete] = useState(false)
+  const [deleting, setDeleting] = useState(false)
+  const [togglingStatus, setTogglingStatus] = useState(false)
+
+  useEffect(() => {
+    let cancelled = false
+    setLoading(true)
+    setError(null)
+    setSaveResult(null)
+
+    axios.get(`${API_BASE}/jobs/${jobId}`)
+      .then(res => {
+        if (!cancelled) {
+          const d = res.data
+          setJob(d)
+          setEditedText(d.reviewed_text ?? d.ocr_text ?? '')
+          setEditDescribeText(d.describe_text ?? '')
+          setEditFreeformText(d.freeform_text ?? '')
+          setEditAuthor(d.author || '')
+          setEditBook(d.book || '')
+          setEditChapter(d.chapter || '')
+          setEditPage(d.page || '')
+          setReviewerName(d.reviewer_name || '')
+          // Default to the OCR tab when there's OCR text, otherwise Description
+          if (d.reviewed_text || d.ocr_text) setActiveTab('ocr')
+          else setActiveTab('describe')
+        }
+      })
+      .catch(err => {
+        if (!cancelled) setError(err.response?.data?.detail || err.message)
+      })
+      .finally(() => { if (!cancelled) setLoading(false) })
+
+    return () => { cancelled = true }
+  }, [jobId])
+
+  // Default the Describe model to the job's original model (if available) or the registry default
+  useEffect(() => {
+    if (!describeModel && models.length > 0) {
+      const def = models.find(m => m.default) || models[0]
+      const fromJob = job?.ocr_model && models.some(m => m.id === job.ocr_model) ? job.ocr_model : null
+      setDescribeModel(fromJob || def.id)
+    }
+  }, [models, job, describeModel])
+
+  const handleGenerateDescribe = async () => {
+    setGeneratingDescribe(true)
+    setSaveResult(null)
+    try {
+      const res = await axios.post(`${API_BASE}/jobs/${jobId}/describe`, {
+        model: describeModel || null,
+      })
+      setJob(res.data)
+      setEditDescribeText(res.data.describe_text || '')
+      onReviewed(res.data)
+    } catch (err) {
+      setSaveResult({ success: false, error: err.response?.data?.detail || err.message })
+    } finally {
+      setGeneratingDescribe(false)
+    }
+  }
+
+  const handleSave = async () => {
+    if (!reviewerName.trim()) {
+      setSaveResult({ success: false, error: 'Reviewer name is required.' })
+      return
+    }
+    setSubmitting(true)
+    setSaveResult(null)
+    try {
+      const res = await axios.put(`${API_BASE}/jobs/${jobId}/review`, {
+        reviewed_text: editedText,
+        reviewer_name: reviewerName.trim(),
+        author: editAuthor,
+        book: editBook,
+        chapter: editChapter,
+        page: editPage,
+        describe_text: editDescribeText || null,
+        freeform_text: editFreeformText || null,
+      })
+      setJob(res.data)
+      setSaveResult({ success: true })
+      onReviewed(res.data)
+    } catch (err) {
+      setSaveResult({ success: false, error: err.response?.data?.detail || err.message })
+    } finally {
+      setSubmitting(false)
+    }
+  }
+
+  const handleToggleStatus = async () => {
+    // Marking reviewed accepts BOTH the reviewed document text and the description,
+    // so it goes through the full review save (not a status-only flip).
+    if (!isReviewed) {
+      setTogglingStatus(true)
+      try {
+        await handleSave()
+      } finally {
+        setTogglingStatus(false)
+      }
+      return
+    }
+
+    // Reverting to unreviewed preserves the saved reviewed text and description.
+    setTogglingStatus(true)
+    setSaveResult(null)
+    try {
+      const res = await axios.put(`${API_BASE}/jobs/${jobId}/status`, {
+        status: 'unreviewed',
+        reviewer_name: reviewerName.trim() || null,
+      })
+      setJob(res.data)
+      setReviewerName(res.data.reviewer_name || '')
+      onReviewed(res.data)
+    } catch (err) {
+      setSaveResult({ success: false, error: err.response?.data?.detail || err.message })
+    } finally {
+      setTogglingStatus(false)
+    }
+  }
+
+  const handleDelete = async () => {
+    setDeleting(true)
+    try {
+      await axios.delete(`${API_BASE}/jobs/${jobId}`)
+      onDeleted(jobId)
+    } catch (err) {
+      setSaveResult({ success: false, error: err.response?.data?.detail || err.message })
+      setConfirmDelete(false)
+    } finally {
+      setDeleting(false)
+    }
+  }
+
+  const isReviewed = job?.status === 'reviewed'
+
+  return (
+    <motion.div
+      key={jobId}
+      initial={{ opacity: 0, y: 20 }}
+      animate={{ opacity: 1, y: 0 }}
+      exit={{ opacity: 0, y: -20 }}
+      className="flex flex-col gap-4"
+    >
+      {/* Top bar */}
+      <div className="flex items-center gap-4 flex-shrink-0">
+        <motion.button
+          onClick={onClose}
+          className="flex items-center gap-2 glass glass-hover px-4 py-2 rounded-xl text-sm text-gray-300"
+          whileHover={{ scale: 1.02 }} whileTap={{ scale: 0.98 }}
+        >
+          <ChevronLeft className="w-4 h-4" />
+          Back to results
+        </motion.button>
+        {job && (
+          <>
+            <StatusBadge status={job.status} />
+            <motion.button
+              onClick={handleToggleStatus}
+              disabled={togglingStatus}
+              title={isReviewed ? 'Revert to unreviewed' : 'Mark as reviewed'}
+              className={`flex items-center gap-1 px-3 py-1.5 rounded-lg text-xs font-medium transition-colors disabled:opacity-50 ${
+                isReviewed
+                  ? 'glass glass-hover text-amber-400 hover:bg-amber-500/10'
+                  : 'glass glass-hover text-green-400 hover:bg-green-500/10'
+              }`}
+              whileHover={!togglingStatus ? { scale: 1.02 } : {}}
+              whileTap={!togglingStatus ? { scale: 0.98 } : {}}
+            >
+              {togglingStatus ? (
+                <Loader2 className="w-3.5 h-3.5 animate-spin" />
+              ) : isReviewed ? (
+                <Clock className="w-3.5 h-3.5" />
+              ) : (
+                <CheckCircle2 className="w-3.5 h-3.5" />
+              )}
+              {isReviewed ? 'Mark Unreviewed' : 'Mark Reviewed'}
+            </motion.button>
+            <span className="text-xs text-gray-500 font-mono hidden sm:block">{job.id}</span>
+          </>
+        )}
+        <div className="ml-auto flex items-center gap-2">
+          {confirmDelete ? (
+            <>
+              <span className="text-xs text-red-400">Delete this job permanently?</span>
+              <motion.button
+                onClick={handleDelete}
+                disabled={deleting}
+                className="flex items-center gap-1 px-3 py-2 rounded-xl text-sm font-medium bg-red-600 hover:bg-red-500 disabled:opacity-50"
+                whileHover={{ scale: 1.02 }} whileTap={{ scale: 0.98 }}
+              >
+                {deleting ? <Loader2 className="w-4 h-4 animate-spin" /> : <Trash2 className="w-4 h-4" />}
+                Confirm
+              </motion.button>
+              <motion.button
+                onClick={() => setConfirmDelete(false)}
+                className="px-3 py-2 rounded-xl text-sm glass glass-hover text-gray-300"
+                whileHover={{ scale: 1.02 }} whileTap={{ scale: 0.98 }}
+              >
+                Cancel
+              </motion.button>
+            </>
+          ) : (
+            <motion.button
+              onClick={() => setConfirmDelete(true)}
+              className="flex items-center gap-2 px-3 py-2 rounded-xl text-sm glass glass-hover text-red-400 hover:bg-red-500/10"
+              whileHover={{ scale: 1.02 }} whileTap={{ scale: 0.98 }}
+            >
+              <Trash2 className="w-4 h-4" />
+              Delete
+            </motion.button>
+          )}
+        </div>
+      </div>
+
+      {loading && (
+        <div className="flex-1 flex items-center justify-center">
+          <Loader2 className="w-8 h-8 animate-spin text-purple-400" />
+        </div>
+      )}
+
+      {error && (
+        <div className="glass p-4 rounded-xl border-red-500/30 bg-red-500/10 flex-shrink-0">
+          <p className="text-sm text-red-400">{error}</p>
+        </div>
+      )}
+
+      {job && !loading && (
+        <>
+          {/* Image + Text */}
+          <div className="grid gap-6" style={{ gridTemplateColumns: '1fr 1fr', height: '130vh' }}>
+            <div className="glass rounded-2xl overflow-hidden flex items-center justify-center bg-black/20 h-full">
+              <img
+                src={`${API_BASE}/jobs/${job.id}/image`}
+                alt="Job source"
+                className="w-full h-full object-contain"
+                onError={e => { e.target.style.display = 'none' }}
+              />
+            </div>
+            <div className="glass rounded-2xl p-4 flex flex-col h-full">
+              {/* Tabs — only show tabs that have content */}
+              {(() => {
+                const tabs = [
+                  job.ocr_text || job.reviewed_text ? { id: 'ocr', label: 'OCR Text' } : null,
+                  { id: 'describe', label: 'Description' },
+                ].filter(Boolean)
+                return tabs.length > 1 ? (
+                  <div className="flex gap-1 mb-3 flex-shrink-0">
+                    {tabs.map(t => (
+                      <button
+                        key={t.id}
+                        onClick={() => setActiveTab(t.id)}
+                        className={`px-3 py-1 rounded-lg text-xs font-medium transition-colors ${
+                          activeTab === t.id
+                            ? 'bg-purple-600 text-white'
+                            : 'bg-white/5 text-gray-400 hover:bg-white/10'
+                        }`}
+                      >
+                        {t.label}
+                      </button>
+                    ))}
+                  </div>
+                ) : null
+              })()}
+
+              <p className="text-xs text-gray-400 mb-2 flex-shrink-0">
+                {{ ocr: isReviewed ? 'Reviewed Text' : 'OCR Text', describe: 'Description' }[activeTab]}
+                <span className="text-purple-400 ml-1">(editable)</span>
+              </p>
+
+              {activeTab === 'ocr' && (
+                <>
+                  <textarea
+                    value={editedText}
+                    onChange={e => setEditedText(e.target.value)}
+                    className="flex-1 w-full bg-transparent text-sm text-gray-200 font-mono resize-none focus:outline-none min-h-0"
+                    placeholder="OCR text..."
+                  />
+                  {isReviewed && job.ocr_text && (
+                    <details className="flex-shrink-0 mt-2 border-t border-white/10 pt-2">
+                      <summary className="cursor-pointer text-xs text-gray-500 hover:text-gray-400 transition-colors">
+                        Original OCR Text
+                      </summary>
+                      <pre className="text-xs text-gray-600 whitespace-pre-wrap font-mono mt-1 max-h-28 overflow-y-auto">
+                        {job.ocr_text}
+                      </pre>
+                    </details>
+                  )}
+                </>
+              )}
+              {activeTab === 'describe' && (
+                <>
+                  <div className="flex items-center gap-2 mb-2 flex-shrink-0">
+                    <select
+                      value={describeModel}
+                      onChange={e => setDescribeModel(e.target.value)}
+                      disabled={generatingDescribe || models.length === 0}
+                      className="bg-white/5 border border-white/10 rounded-lg px-2 py-1.5 text-xs text-gray-200 focus:outline-none focus:border-purple-500/50"
+                    >
+                      {models.length === 0 && <option value="">No models</option>}
+                      {models.map(m => (
+                        <option key={m.id} value={m.id}>{m.label}{m.default ? ' (default)' : ''}</option>
+                      ))}
+                    </select>
+                    <motion.button
+                      onClick={handleGenerateDescribe}
+                      disabled={generatingDescribe || !describeModel}
+                      className={`flex items-center gap-1.5 px-3 py-1.5 rounded-lg text-xs font-medium transition-all ${
+                        generatingDescribe || !describeModel
+                          ? 'opacity-50 cursor-not-allowed bg-white/5'
+                          : 'bg-gradient-to-r from-violet-600 to-purple-600 hover:from-violet-500 hover:to-purple-500'
+                      }`}
+                      whileHover={!generatingDescribe && describeModel ? { scale: 1.02 } : {}}
+                      whileTap={!generatingDescribe && describeModel ? { scale: 0.98 } : {}}
+                      title="Run Describe on this job's image and save it"
+                    >
+                      {generatingDescribe
+                        ? <><Loader2 className="w-3.5 h-3.5 animate-spin" /> Generating…</>
+                        : <><Sparkles className="w-3.5 h-3.5" /> Generate Description</>}
+                    </motion.button>
+                  </div>
+                  <textarea
+                    value={editDescribeText}
+                    onChange={e => setEditDescribeText(e.target.value)}
+                    className="flex-1 w-full bg-transparent text-sm text-gray-200 font-mono resize-none focus:outline-none min-h-0"
+                    placeholder="No description yet — pick a model and click Generate Description, or type one here."
+                  />
+                </>
+              )}
+            </div>
+          </div>
+
+          {/* Metadata + reviewer row */}
+          <div className="glass p-4 rounded-2xl flex-shrink-0">
+            <datalist id="jd-authors">
+              {(suggestions.authors || []).map(a => <option key={a} value={a} />)}
+            </datalist>
+            <datalist id="jd-books">
+              {(suggestions.books || []).map(b => <option key={b} value={b} />)}
+            </datalist>
+            <datalist id="jd-chapters">
+              {(suggestions.chapters || []).map(c => <option key={c} value={c} />)}
+            </datalist>
+            <datalist id="jd-reviewers">
+              {(suggestions.reviewers || []).map(r => <option key={r} value={r} />)}
+            </datalist>
+            <div className="grid grid-cols-6 gap-4">
+              <div>
+                <label className="text-xs text-gray-400 mb-1 block">Author</label>
+                <input type="text" list="jd-authors" value={editAuthor} onChange={e => setEditAuthor(e.target.value)} placeholder="Author" className={INPUT_CLASS} />
+              </div>
+              <div>
+                <label className="text-xs text-gray-400 mb-1 block">Book</label>
+                <input type="text" list="jd-books" value={editBook} onChange={e => setEditBook(e.target.value)} placeholder="Book title" className={INPUT_CLASS} />
+              </div>
+              <div>
+                <label className="text-xs text-gray-400 mb-1 block">Chapter</label>
+                <input type="text" list="jd-chapters" value={editChapter} onChange={e => setEditChapter(e.target.value)} placeholder="Chapter" className={INPUT_CLASS} />
+              </div>
+              <div>
+                <label className="text-xs text-gray-400 mb-1 block">Page</label>
+                <input type="text" value={editPage} onChange={e => setEditPage(e.target.value)} placeholder="Page" className={INPUT_CLASS} />
+              </div>
+              <div>
+                <label className="text-xs text-gray-400 mb-1 block">Reviewer</label>
+                <input type="text" list="jd-reviewers" value={reviewerName} onChange={e => setReviewerName(e.target.value)} placeholder="Your name" className={INPUT_CLASS} />
+              </div>
+              <div className="flex flex-col justify-end">
+                <motion.button
+                  onClick={handleSave}
+                  disabled={submitting || !reviewerName.trim()}
+                  className={`w-full flex items-center justify-center gap-2 px-4 py-2 rounded-lg font-medium text-sm transition-all ${
+                    submitting || !reviewerName.trim()
+                      ? 'opacity-50 cursor-not-allowed bg-white/5'
+                      : isReviewed
+                        ? 'bg-gradient-to-r from-blue-600 to-indigo-600 hover:from-blue-500 hover:to-indigo-500'
+                        : 'bg-gradient-to-r from-green-600 to-emerald-600 hover:from-green-500 hover:to-emerald-500'
+                  }`}
+                  whileHover={!submitting && reviewerName.trim() ? { scale: 1.02 } : {}}
+                  whileTap={!submitting && reviewerName.trim() ? { scale: 0.98 } : {}}
+                >
+                  {submitting ? (
+                    <><Loader2 className="w-4 h-4 animate-spin" /> Saving...</>
+                  ) : isReviewed ? (
+                    <><Save className="w-4 h-4" /> Save Changes</>
+                  ) : (
+                    <><CheckCircle2 className="w-4 h-4" /> Mark Reviewed</>
+                  )}
+                </motion.button>
+              </div>
+            </div>
+
+            {!isReviewed && (
+              <p className="text-xs text-gray-500 mt-2">
+                Marking reviewed accepts both the reviewed document text and the description.
+              </p>
+            )}
+
+            {saveResult && (
+              <motion.div
+                initial={{ opacity: 0, y: -4 }} animate={{ opacity: 1, y: 0 }}
+                className={`mt-3 p-2 rounded-lg text-xs ${saveResult.success ? 'bg-green-500/10 text-green-400' : 'bg-red-500/10 text-red-400'}`}
+              >
+                {saveResult.success
+                  ? (isReviewed ? 'Changes saved!' : 'Job marked as reviewed!')
+                  : saveResult.error}
+              </motion.div>
+            )}
+
+            {/* Read-only info row */}
+            <div className="flex gap-6 mt-3 pt-3 border-t border-white/10">
+              {job.submitted_at && (
+                <span className="text-xs text-gray-500">Submitted: {new Date(job.submitted_at).toLocaleString()}</span>
+              )}
+              {isReviewed && job.reviewed_at && (
+                <span className="text-xs text-gray-500">Last reviewed: {new Date(job.reviewed_at).toLocaleString()}</span>
+              )}
+              {job.mode && <span className="text-xs text-gray-500">Mode: {job.mode}</span>}
+              {job.ocr_model && <span className="text-xs text-gray-500">Model: {job.ocr_model}</span>}
+            </div>
+          </div>
+        </>
+      )}
+    </motion.div>
+  )
+}
+
+// ─────────────────────────────────────────────────────────────
+// Search / List view
+// ─────────────────────────────────────────────────────────────
+export default function JobsPanel() {
+  const suggestions = useSuggestions()
+  const [search, setSearch] = useState('')
+  const [filterStatus, setFilterStatus] = useState('')
+  const [filterAuthor, setFilterAuthor] = useState('')
+  const [filterBook, setFilterBook] = useState('')
+  const [jobs, setJobs] = useState([])
+  const [total, setTotal] = useState(0)
+  const [page, setPage] = useState(0)
+  const [loading, setLoading] = useState(false)
+  const [error, setError] = useState(null)
+  const [selectedJobId, setSelectedJobId] = useState(null)
+
+  const LIMIT = 20
+
+  const fetchJobs = useCallback(async (pageNum = 0) => {
+    setLoading(true)
+    setError(null)
+    try {
+      const params = new URLSearchParams()
+      if (search.trim()) params.set('search', search.trim())
+      if (filterStatus) params.set('status', filterStatus)
+      if (filterAuthor.trim()) params.set('author', filterAuthor.trim())
+      if (filterBook.trim()) params.set('book', filterBook.trim())
+      params.set('limit', LIMIT)
+      params.set('offset', pageNum * LIMIT)
+
+      const res = await axios.get(`${API_BASE}/jobs?${params}`)
+      setJobs(res.data.jobs)
+      setTotal(res.data.total)
+      setPage(pageNum)
+    } catch (err) {
+      setError(err.response?.data?.detail || err.message)
+    } finally {
+      setLoading(false)
+    }
+  }, [search, filterStatus, filterAuthor, filterBook])
+
+  useEffect(() => { fetchJobs(0) }, []) // eslint-disable-line react-hooks/exhaustive-deps
+
+  const handleReviewed = (updatedJob) => {
+    setJobs(prev => prev.map(j => j.id === updatedJob.id ? { ...j, ...updatedJob } : j))
+  }
+
+  const totalPages = Math.ceil(total / LIMIT)
+
+  // When a job is selected show full-screen detail
+  if (selectedJobId) {
+    return (
+      <AnimatePresence mode="wait">
+        <JobDetail
+          key={selectedJobId}
+          jobId={selectedJobId}
+          onClose={() => setSelectedJobId(null)}
+          onReviewed={handleReviewed}
+          onDeleted={(id) => {
+            setJobs(prev => prev.filter(j => j.id !== id))
+            setTotal(prev => prev - 1)
+            setSelectedJobId(null)
+          }}
+          suggestions={suggestions}
+        />
+      </AnimatePresence>
+    )
+  }
+
+  return (
+    <motion.div
+      key="job_list"
+      initial={{ opacity: 0, y: 20 }}
+      animate={{ opacity: 1, y: 0 }}
+      exit={{ opacity: 0, y: -20 }}
+      className="space-y-4"
+    >
+      {/* Search form */}
+      <div className="glass p-4 rounded-2xl space-y-3">
+        <form onSubmit={e => { e.preventDefault(); fetchJobs(0) }} className="flex gap-2">
+          <input
+            type="text"
+            value={search}
+            onChange={e => setSearch(e.target.value)}
+            placeholder="Search all fields..."
+            className={`${INPUT_CLASS} flex-1`}
+          />
+          <motion.button
+            type="submit"
+            className="flex items-center gap-2 px-4 py-2 rounded-lg bg-gradient-to-r from-purple-600 to-cyan-600 text-sm font-medium"
+            whileHover={{ scale: 1.02 }} whileTap={{ scale: 0.98 }}
+          >
+            <Search className="w-4 h-4" /> Search
+          </motion.button>
+        </form>
+
+        <datalist id="jp-authors">
+          {suggestions.authors.map(a => <option key={a} value={a} />)}
+        </datalist>
+        <datalist id="jp-books">
+          {(suggestions.books || []).map(b => <option key={b} value={b} />)}
+        </datalist>
+        <div className="grid grid-cols-3 gap-2">
+          <select value={filterStatus} onChange={e => setFilterStatus(e.target.value)} className={INPUT_CLASS}>
+            <option value="">All statuses</option>
+            <option value="unreviewed">Unreviewed</option>
+            <option value="reviewed">Reviewed</option>
+          </select>
+          <input type="text" list="jp-authors" value={filterAuthor} onChange={e => setFilterAuthor(e.target.value)} placeholder="Author..." className={INPUT_CLASS} />
+          <input type="text" list="jp-books" value={filterBook} onChange={e => setFilterBook(e.target.value)} placeholder="Book..." className={INPUT_CLASS} />
+        </div>
+
+        <div className="flex items-center justify-between">
+          <span className="text-xs text-gray-500">{total} job{total !== 1 ? 's' : ''} found</span>
+          <button onClick={() => fetchJobs(page)} className="flex items-center gap-1 text-xs text-gray-400 hover:text-gray-200 transition-colors">
+            <RefreshCw className="w-3 h-3" /> Refresh
+          </button>
+        </div>
+      </div>
+
+      {loading && <div className="flex justify-center py-8"><Loader2 className="w-6 h-6 animate-spin text-purple-400" /></div>}
+
+      {error && (
+        <div className="glass p-4 rounded-xl border-red-500/30 bg-red-500/10">
+          <p className="text-sm text-red-400">{error}</p>
+        </div>
+      )}
+
+      {!loading && !error && jobs.length === 0 && (
+        <div className="glass p-8 rounded-2xl text-center">
+          <FileText className="w-10 h-10 mx-auto mb-3 text-gray-600" />
+          <p className="text-gray-400">No jobs found</p>
+          <p className="text-xs text-gray-500 mt-1">Commit your first OCR job from the New Job tab</p>
+        </div>
+      )}
+
+      {/* Results grid */}
+      <div className="grid grid-cols-1 sm:grid-cols-2 lg:grid-cols-3 xl:grid-cols-4 gap-3">
+        <AnimatePresence>
+          {jobs.map(job => (
+            <motion.button
+              key={job.id}
+              onClick={() => setSelectedJobId(job.id)}
+              className="text-left glass p-4 rounded-xl border border-white/5 hover:border-white/20 hover:bg-white/5 transition-all"
+              initial={{ opacity: 0, y: 10 }}
+              animate={{ opacity: 1, y: 0 }}
+              exit={{ opacity: 0 }}
+              whileHover={{ scale: 1.02 }}
+              whileTap={{ scale: 0.98 }}
+              layout
+            >
+              <div className="flex items-start justify-between gap-2 mb-2">
+                <StatusBadge status={job.status} />
+              </div>
+              {job.book && <p className="text-sm font-medium text-gray-200 truncate">{job.book}</p>}
+              <div className="flex items-center gap-2 mt-0.5">
+                {job.chapter && <span className="text-xs text-gray-500">Ch. {job.chapter}</span>}
+                {job.page && <span className="text-xs text-gray-500">p. {job.page}</span>}
+              </div>
+              {job.author && <p className="text-xs text-gray-400 mt-1">{job.author}</p>}
+              <div className="flex items-center justify-between mt-2">
+                <p className="text-xs text-gray-600 font-mono">{new Date(job.submitted_at).toLocaleDateString()}</p>
+                {job.ocr_model && <span className="text-[10px] text-gray-500 truncate ml-2">{job.ocr_model}</span>}
+              </div>
+            </motion.button>
+          ))}
+        </AnimatePresence>
+      </div>
+
+      {totalPages > 1 && (
+        <div className="flex items-center justify-center gap-3">
+          <button onClick={() => fetchJobs(page - 1)} disabled={page === 0} className="glass glass-hover p-2 rounded-lg disabled:opacity-30">
+            <ChevronLeft className="w-4 h-4" />
+          </button>
+          <span className="text-sm text-gray-400">Page {page + 1} of {totalPages}</span>
+          <button onClick={() => fetchJobs(page + 1)} disabled={page >= totalPages - 1} className="glass glass-hover p-2 rounded-lg disabled:opacity-30">
+            <ChevronRight className="w-4 h-4" />
+          </button>
+        </div>
+      )}
+    </motion.div>
+  )
+}

frontend/src/components/MetadataForm.jsx

@@ -0,0 +1,77 @@
+import { BookOpen } from 'lucide-react'
+
+export default function MetadataForm({ metadata, onChange, suggestions = {} }) {
+  const { author, book, chapter, page } = metadata
+  const { authors = [], books = [], chapters = [] } = suggestions
+
+  const field = (key) => (e) => onChange({ ...metadata, [key]: e.target.value })
+
+  const inputClass =
+    'w-full bg-white/5 border border-white/10 rounded-lg px-3 py-2 text-sm text-gray-200 ' +
+    'placeholder-gray-600 focus:outline-none focus:border-purple-500/50 transition-colors'
+
+  return (
+    <div className="glass p-4 rounded-2xl space-y-3">
+      <div className="flex items-center gap-2">
+        <BookOpen className="w-4 h-4 text-purple-400" />
+        <h3 className="text-sm font-medium text-gray-300">Job Metadata</h3>
+      </div>
+
+      <datalist id="mf-authors">
+        {authors.map(a => <option key={a} value={a} />)}
+      </datalist>
+      <datalist id="mf-books">
+        {books.map(b => <option key={b} value={b} />)}
+      </datalist>
+      <datalist id="mf-chapters">
+        {chapters.map(c => <option key={c} value={c} />)}
+      </datalist>
+
+      <div className="grid grid-cols-2 gap-3">
+        <div>
+          <label className="text-xs text-gray-400 mb-1 block">Author</label>
+          <input
+            type="text"
+            list="mf-authors"
+            value={author}
+            onChange={field('author')}
+            placeholder="Author name"
+            className={inputClass}
+          />
+        </div>
+        <div>
+          <label className="text-xs text-gray-400 mb-1 block">Book</label>
+          <input
+            type="text"
+            list="mf-books"
+            value={book}
+            onChange={field('book')}
+            placeholder="Book title"
+            className={inputClass}
+          />
+        </div>
+        <div>
+          <label className="text-xs text-gray-400 mb-1 block">Chapter</label>
+          <input
+            type="text"
+            list="mf-chapters"
+            value={chapter}
+            onChange={field('chapter')}
+            placeholder="Chapter"
+            className={inputClass}
+          />
+        </div>
+        <div>
+          <label className="text-xs text-gray-400 mb-1 block">Page</label>
+          <input
+            type="text"
+            value={page}
+            onChange={field('page')}
+            placeholder="Page number"
+            className={inputClass}
+          />
+        </div>
+      </div>
+    </div>
+  )
+}

frontend/src/components/ModeSelector.jsx

@@ -1,29 +1,17 @@
 import { motion } from 'framer-motion'
-import { FileText, Eye, Search, Wand2 } from 'lucide-react'
+import { FileText, Eye } from 'lucide-react'
 
 const modes = [
-  { id: 'plain_ocr', name: 'Plain OCR', icon: FileText, color: 'from-blue-500 to-cyan-500', desc: 'Extract raw text', needsInput: false },
+  { id: 'plain_ocr', name: 'Plain OCR', icon: FileText, color: 'from-blue-500 to-cyan-500', desc: 'Extract raw text' },
-  { id: 'describe', name: 'Describe', icon: Eye, color: 'from-violet-500 to-purple-500', desc: 'Image description', needsInput: false },
+  { id: 'describe', name: 'Describe', icon: Eye, color: 'from-violet-500 to-purple-500', desc: 'Image description' },
-  { id: 'find_ref', name: 'Find', icon: Search, color: 'from-yellow-500 to-orange-500', desc: 'Locate specific terms', needsInput: 'findTerm' },
-  { id: 'freeform', name: 'Freeform', icon: Wand2, color: 'from-fuchsia-500 to-pink-500', desc: 'Custom prompt', needsInput: 'prompt' },
 ]
 
-export default function ModeSelector({ 
+export default function ModeSelector({ mode, onModeChange }) {
-  mode, 
-  onModeChange, 
-  prompt, 
-  onPromptChange,
-  findTerm,
-  onFindTermChange
-}) {
-  const selectedMode = modes.find(m => m.id === mode)
-  const needsInput = selectedMode?.needsInput
-
   return (
     <div className="glass p-4 rounded-2xl space-y-3">
       <h3 className="text-sm font-semibold text-gray-200">Mode</h3>
 
-      <div className="grid grid-cols-4 gap-2">
+      <div className="grid grid-cols-2 gap-2">
         {modes.map((m) => {
           const Icon = m.icon
           const isSelected = mode === m.id
@@ -32,6 +20,7 @@ export default function ModeSelector({
             <motion.button
               key={m.id}
               onClick={() => onModeChange(m.id)}
+              title={m.desc}
               className={`
                 relative p-2 rounded-xl text-center transition-all
                 ${isSelected
@@ -68,38 +57,6 @@ export default function ModeSelector({
           )
         })}
       </div>
-
-      {needsInput === 'findTerm' && (
-        <motion.div
-          initial={{ opacity: 0, height: 0 }}
-          animate={{ opacity: 1, height: 'auto' }}
-          exit={{ opacity: 0, height: 0 }}
-        >
-          <input
-            type="text"
-            value={findTerm}
-            onChange={(e) => onFindTermChange(e.target.value)}
-            placeholder="Enter term to find (e.g., Total, Invoice #)"
-            className="w-full bg-white/5 border border-white/10 rounded-xl px-3 py-2 text-sm focus:outline-none focus:border-purple-500 transition-colors"
-          />
-        </motion.div>
-      )}
-
-      {needsInput === 'prompt' && (
-        <motion.div
-          initial={{ opacity: 0, height: 0 }}
-          animate={{ opacity: 1, height: 'auto' }}
-          exit={{ opacity: 0, height: 0 }}
-        >
-          <textarea
-            value={prompt}
-            onChange={(e) => onPromptChange(e.target.value)}
-            placeholder="Enter your custom prompt..."
-            className="w-full bg-white/5 border border-white/10 rounded-xl px-3 py-2 text-sm focus:outline-none focus:border-purple-500 transition-colors resize-none"
-            rows={2}
-          />
-        </motion.div>
-      )}
     </div>
   )
 }

frontend/src/components/ModelSelector.jsx

@@ -0,0 +1,33 @@
+import { Cpu } from 'lucide-react'
+
+const SELECT_CLASS =
+  'w-full bg-white/5 border border-white/10 rounded-lg px-3 py-2 text-sm text-gray-200 ' +
+  'focus:outline-none focus:border-purple-500/50 transition-colors'
+
+// Dropdown to pick which OCR model runs the analysis.
+// `models` comes from the useModels() hook; `value` is the selected model id.
+export default function ModelSelector({ models, value, onChange, loading }) {
+  return (
+    <div className="glass p-4 rounded-2xl space-y-3">
+      <div className="flex items-center gap-2">
+        <Cpu className="w-4 h-4 text-purple-400" />
+        <h3 className="text-sm font-semibold text-gray-200">Model</h3>
+      </div>
+
+      <select
+        value={value || ''}
+        onChange={e => onChange(e.target.value)}
+        disabled={loading || models.length === 0}
+        className={SELECT_CLASS}
+      >
+        {loading && <option value="">Loading models…</option>}
+        {!loading && models.length === 0 && <option value="">No models available</option>}
+        {models.map(m => (
+          <option key={m.id} value={m.id}>
+            {m.label}{m.default ? ' (default)' : ''}
+          </option>
+        ))}
+      </select>
+    </div>
+  )
+}

frontend/src/components/PDFProcessor.jsx

@@ -5,7 +5,7 @@ import axios from 'axios'
 
 const API_BASE = import.meta.env.VITE_API_URL || '/api'
 
-function PDFProcessor({ pdfFile, mode, prompt, advancedSettings, includeCaption }) {
+function PDFProcessor({ pdfFile, mode, prompt, model, advancedSettings, includeCaption }) {
   const [processing, setProcessing] = useState(false)
   const [progress, setProgress] = useState(0)
   const [result, setResult] = useState(null)
@@ -29,6 +29,7 @@ function PDFProcessor({ pdfFile, mode, prompt, advancedSettings, includeCaption
     try {
       const formData = new FormData()
       formData.append('pdf_file', pdfFile)
+      if (model) formData.append('model', model)
       formData.append('mode', mode)
       formData.append('prompt', prompt)
       formData.append('output_format', outputFormat)
@@ -80,7 +81,7 @@ function PDFProcessor({ pdfFile, mode, prompt, advancedSettings, includeCaption
     } finally {
       setProcessing(false)
     }
-  }, [pdfFile, mode, prompt, outputFormat, includeCaption, advancedSettings])
+  }, [pdfFile, mode, prompt, model, outputFormat, includeCaption, advancedSettings])
 
   const handleDownloadJSON = useCallback(() => {
     if (!result || outputFormat !== 'json') return

frontend/src/components/ResultPanel.jsx

@@ -205,7 +205,7 @@ export default function ResultPanel({ result, loading, imagePreview, onCopy, onD
             exit={{ opacity: 0, y: -20 }}
             className="space-y-4"
           >
-            {/* Preview with boxes */}
+            {/* Preview with boxes (grounding modes) */}
             {imagePreview && result.boxes && result.boxes.length > 0 && (
               <div className="relative rounded-xl overflow-hidden border border-white/10 bg-black">
                 <img
@@ -226,15 +226,13 @@ export default function ResultPanel({ result, loading, imagePreview, onCopy, onD
               </div>
             )}
 
-            {/* Text result */}
+            {/* Rendered text result */}
             <div className="bg-white/5 border border-white/10 rounded-xl p-4 max-h-96 overflow-y-auto">
               {isHTML ? (
                 <div
                   className="prose prose-invert prose-sm max-w-none"
                   dangerouslySetInnerHTML={{ __html: DOMPurify.sanitize(result.text) }}
-                  style={{
+                  style={{ color: '#e5e7eb' }}
-                    color: '#e5e7eb',
-                  }}
                 />
               ) : isMarkdown ? (
                 <div className="prose prose-invert prose-sm max-w-none">

frontend/src/hooks/useModels.js

@@ -0,0 +1,24 @@
+import { useState, useEffect } from 'react'
+
+const API_BASE = import.meta.env.VITE_API_URL || '/api'
+
+// Fetches the OCR models available for selection. Returns { models, loading }.
+// Each model: { id, label, capabilities: { grounding, advanced_settings }, default }
+export function useModels() {
+  const [models, setModels] = useState([])
+  const [loading, setLoading] = useState(true)
+
+  useEffect(() => {
+    let cancelled = false
+    fetch(`${API_BASE}/models`)
+      .then(r => (r.ok ? r.json() : null))
+      .then(data => {
+        if (!cancelled && data?.models) setModels(data.models)
+      })
+      .catch(() => {})
+      .finally(() => { if (!cancelled) setLoading(false) })
+    return () => { cancelled = true }
+  }, [])
+
+  return { models, loading }
+}

frontend/src/hooks/useSuggestions.js

@@ -0,0 +1,16 @@
+import { useState, useEffect } from 'react'
+
+const API_BASE = import.meta.env.VITE_API_URL || '/api'
+
+export function useSuggestions() {
+  const [suggestions, setSuggestions] = useState({ authors: [], books: [], chapters: [], reviewers: [] })
+
+  useEffect(() => {
+    fetch(`${API_BASE}/jobs/suggestions`)
+      .then(r => r.ok ? r.json() : null)
+      .then(data => { if (data) setSuggestions(data) })
+      .catch(() => {})
+  }, [])
+
+  return suggestions
+}