Prompt Modes Integration:
- Create local_runtime.py with DOTS_PROMPT_MAP
- Map OutputMode to native dots.ocr prompt modes (prompt_layout_all_en, prompt_ocr, etc.)
- Support dict_promptmode_to_prompt from dots.ocr with fallback prompts
- Add layout_only and region modes to OutputMode enum
2-Stage Q&A Pipeline:
- Create qa_builder.py for 2-stage qa_pairs generation
- Stage 1: PARSER (dots.ocr) → raw JSON via prompt_layout_all_en
- Stage 2: LLM (DAGI Router) → Q&A pairs via mode=qa_build
- Update endpoints.py to use 2-stage pipeline for qa_pairs mode
- Add ROUTER_BASE_URL and ROUTER_TIMEOUT to config
Updates:
- Update inference.py to use local_runtime with native prompts
- Update ollama_client.py to use same prompt map
- Add PROMPT_MODES.md documentation
Model Loader:
- Update model_loader.py with complete dots.ocr loading code
- Proper device detection (CUDA/CPU/MPS) with fallback
- Memory optimization (low_cpu_mem_usage)
- Better error handling and logging
- Support for local model paths and HF Hub
Docker:
- Multi-stage Dockerfile (CPU/CUDA builds)
- docker-compose.yml for parser-service
- .dockerignore for clean builds
- Model cache volume for persistence
Configuration:
- Support DOTS_OCR_MODEL_ID and DEVICE env vars (backward compatible)
- Better defaults and environment variable handling
Deployment:
- Add DEPLOYMENT.md with detailed instructions
- Local deployment (venv)
- Docker Compose deployment
- Ollama runtime setup
- Troubleshooting guide
Integration:
- Add parser-service to main docker-compose.yml
- Configure volumes and networks
- Health checks and dependencies
Ollama Runtime:
- Add ollama_client.py for Ollama API integration
- Support for dots-ocr model via Ollama
- Add OLLAMA_BASE_URL configuration
- Update inference.py to support Ollama runtime (RUNTIME_TYPE=ollama)
- Update endpoints to handle async Ollama calls
- Alternative to local transformers model
RAG Implementation Plan:
- Create TODO-RAG.md with detailed Haystack integration plan
- Document Store setup (pgvector)
- Embedding model selection
- Ingest pipeline (PARSER → RAG)
- Query pipeline (RAG → LLM)
- Integration with DAGI Router
- Bot commands (/upload_doc, /ask_doc)
- Testing strategy
Now supports three runtime modes:
1. Local transformers (RUNTIME_TYPE=local)
2. Ollama (RUNTIME_TYPE=ollama)
3. Dummy (USE_DUMMY_PARSER=true)
G.2.5 - Tests:
- Add pytest test suite with fixtures
- test_preprocessing.py - PDF/image loading, normalization, validation
- test_postprocessing.py - chunks, QA pairs, markdown generation
- test_inference.py - dummy parser and inference functions
- test_api.py - API endpoint tests
- Add pytest.ini configuration
G.1.3 - dots.ocr Integration:
- Update model_loader.py with real model loading code
- Support for AutoModelForVision2Seq and AutoProcessor
- Device handling (CUDA/CPU/MPS) with fallback
- Error handling with dummy fallback option
- Update inference.py with real model inference
- Process images through model
- Generate and decode outputs
- Parse model output to blocks
- Add model_output_parser.py
- Parse JSON or plain text model output
- Convert to structured blocks
- Layout detection support (placeholder)
Dependencies:
- Add pytest, pytest-asyncio, httpx for testing
