Spaces:
Sleeping
Sleeping
Commit History
18 days ago
-
Support Mistral vision chat completions models and update default config
8394687
-
Clean up dead code and add client CRM integration architecture guide
4124165 -
Implement OCR pipeline enhancements: Pydantic schemas, validation gate, retry, and evaluation harness (excluding images)
497c3ef -
Evolve local prototype to production backend API and decoupled HF Spaces UI
e4e6ceb
-
Cleanup codebase and update documentation for the VLM-direct JSON architecture
a7794c9 -
feat: PaddleOCR-VL BL extraction via TensorRT-LLM (schema-constrained JSON)
Replace the LFM2-VL markdown service with a delivery-note (Bon de Livraison) extractor on PaddleOCR-VL, and move extraction from CPU regex to in-model JSON generation. - bl_extractor.py: Pydantic DeliveryNoteExtractionSchema (raw typed fields), hardened guided-decoding JSON schema, data-only prompt, orjson parsing, TrtLlmBackend (LLM + SamplingParams + GuidedDecodingParams, PagedAttention, continuous batching) and DeliveryNoteExtractor. No regex. - Confidence: data-only payload; one document-level model_confidence from token logprobs (exp(mean logprob)); no per-field self-report. - preprocess.py: speed-first NaViT rules (1MP cap, BILINEAR, 28-patch align, PDF 150 DPI; no deskew/denoise/binarize/pad). - pipeline.py: HFBackend (Path B) keeps API/CLI runnable without a GPU; TrtLlmBackend is the production target on precomputed vision embeddings. - FastAPI async job queue retained; cli.py batch-extracts OCR/images. - Tests: torch-free schema/parse/extractor + API smoke (93 che…e9f136e
-
feat: switch to LFM2-VL-450M, return structured Markdown
9d1f206
-
feat: async job-queue OCR (/predict -> job_id, poll /result)
b794869 -
initial deployment
ada3261