Settings — Extraction strategy

Upload an embedding model (offline)

For locked-down machines that can't reach huggingface.co. On a PC with internet, run:

pip install huggingface_hub
huggingface-cli download ibm-granite/granite-embedding-97m-multilingual-r2 --local-dir granite-97m
cd granite-97m
# Windows PowerShell:
Compress-Archive -Path .\* -DestinationPath ..\granite-97m.zip
# Linux/macOS:
zip -r ../granite-97m.zip .

Drop the resulting .zip below. The model is extracted under data/models/embeddings/ and activated immediately — no network calls.

Drag a model .zip here, or click to browse

Order the providers Inbound should try, top to bottom. For each column, the first provider whose confidence beats its threshold wins; later providers can still fill columns that earlier ones missed. Toggle a provider off to skip it entirely.

Embedding model (used by BGE-small strategy)

The model that builds the per-template embedding index. Smaller / faster ≠ worse for this pipeline — we only embed short token windows. After changing the model, re-run Retrain on each template so its index rebuilds.

Or paste any Hugging Face repo ID or local path:

Examples:

Pure rules: rules only.
Rules → CRF: deterministic floor, CRF fills gaps.
Rules → CRF → BGE-small: add semantic retrieval for hard cases (needs sentence-transformers).
Rules → CRF → Claude: send only the un-confident leftovers to Claude.
Claude only: disable everything else, set Claude min-confidence to 0.