Skip to content

models.md

Latest commit

 

History

History
49 lines (39 loc) · 2.27 KB

models.md

File metadata and controls

49 lines (39 loc) · 2.27 KB
Raw

OCR models

Open Ingest ships four OCR backends. Pick the one that matches your constraints (cost, latency, features, deployment model) and pass it via ocr_model on the ParseRequest.

Backend comparison

ocr_model Provider Native PDF Cell bboxes Forms / KV Barcodes Custom prompt Hardware
dots-ocr DotsOCR + Ovis2.5 converts ✅ ✅ ✅ ✅ CUDA GPU (local or managed)
azure-di (default) Azure Document Intelligence ✅ ✅ ✅ — — Azure cloud
textract AWS Textract ✅ ✅ ✅ — — AWS cloud
gemini Google Gemini VLM ✅ ✅ partial — — Google cloud

dots-ocr is the backend with the most serving work in this repo — open-sourced with the full setup (vLLM, two-stage Ovis figure OCR, masked-region retries). It needs a CUDA-equipped host — either a --local run on your own GPU, or a managed Tensorlake GPU deployment. The @function() decorators are pre-pinned to H100/A100-80GB, but GPU workers aren't part of the open serverless tier today — reach out to support@tensorlake.ai if you'd like a managed deployment. Requests that ask for dots-ocr without a GPU fail at task start with a RequestError suggesting one of the cloud backends.

Picking a model

  • Don't know yet? Start with azure-di — fast cloud OCR with cell-level table bboxes, no GPU required.
  • Need signatures or async S3 jobs? textract.
  • VLM-style semantic OCR? gemini — slower but reads context.
  • Complex documents on your own GPU? dots-ocr ships with the full serving setup (vLLM, two-stage Ovis figure OCR, masked-region retries). Needs a CUDA host (local or managed).

Required env vars

See .env.example — each backend lists its keys with the features it unlocks. Missing keys disable that backend; the rest of the pipeline still runs.

Figure OCR (dots-ocr only)

When ocr_model='dots-ocr', DotsOCR outputs are post-processed by Ovis2.5-9B running on a separate GPU container. The Ovis pass classifies each cropped figure (BARCODE, CHART, DIAGRAM, FORM, TABLE, OTHER) and extracts content with a type-specific prompt.