Skip to content
Draft
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions README.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# DocExtract: document extraction with an eval gate in CI

> **Client work.** I built DocExtract under a paid engagement for a client I can't name. This repository is the version I can share publicly. An authorized walkthrough is available on request.

DocExtract turns invoices, receipts, statements and other PDFs or scans into validated structured records you can search. Every pull request that touches prompts or the extractor re-scores a committed set of extraction outputs, and the check fails if the score falls below a floor or drops from baseline. This offline check scores frozen predictions; it does not rerun the extractor or detect prompt/model regressions. Those require a live evaluation.

Under the hood: FastAPI, a two-pass Claude extraction pipeline, pgvector, and agentic RAG for questions over stored documents.
Expand Down Expand Up @@ -201,8 +203,6 @@ The ADRs live in [docs/adr/](docs/adr/). The key ones:
- There are two separate judges. The in-app judge (Gemini 2.5 Flash, Claude Haiku fallback) is off by default (`llm_judge_enabled = False` in `app/config.py`). The CI judge (`scripts/eval_llm_judge.py`) runs only in the paid live job, which skips without an API key, and uses its default provider (Anthropic) because the workflow passes no `--provider`.
- The two-pass figures in [ADR-0003](docs/adr/0003-two-pass-extraction.md) (trigger rate, improvement, added latency) have no committed run artifact and are not cited as results here.

**Provenance:** built under a paid client engagement; an authorized walkthrough is available on request.

</details>

## Roadmap
Expand Down
Loading