Support Mistral vision chat completions models and update default config

Clean up dead code and add client CRM integration architecture guide

Implement OCR pipeline enhancements: Pydantic schemas, validation gate, retry, and evaluation harness (excluding images)

Add safe __main__ guard with port-in-use detection

Switch to Docker SDK – single uvicorn process, no port conflicts

Upgrade Gradio to 5.13.0 and update dependencies

Pin huggingface_hub version for Gradio 4.44 compatibility

Remove redundant spaces/ folder

Configure combined FastAPI + Gradio server on Python 3.11

Evolve local prototype to production backend API and decoupled HF Spaces UI

fix: patch create_causal_mask to accept cache_position kwarg

fix: patch model's prepare_inputs_for_generation for cache_position=None

fix: centralize v5.x compat patches in app/_compat.py, uncap transformers

fix: pin transformers<5.0.0 and remove all v5 monkeypatches

fix: add compute_default_rope_parameters monkeypatch on nn.Module for custom RotaryEmbedding weight init in transformers v5

fix: add default rope scaling initialization monkeypatch for transformers v5 compatibility

fix: add torch float8 dtype compatibility patch for newer transformers

Cleanup codebase and update documentation for the VLM-direct JSON architecture

feat: PaddleOCR-VL BL extraction via TensorRT-LLM (schema-constrained JSON)

feat: switch to LFM2-VL-450M, return structured Markdown

feat: async job-queue OCR (/predict -> job_id, poll /result)

fix: bf16 on CPU to fit 16GB, use_fast, silence gen warnings

initial deployment

initial commit