v0.1.2

ProofCite v0.1.2 — Hardened Core, Evals Scaffold, Pluggable Retrieval

Highlights

• Tests + CI: Added pytest unit tests (core retrieval, spans, filters) and a GitHub Actions workflow.
• Config: Pydantic Settings (PROOFCITE_ env) for defaults — threshold, rerank, span gap, segmentation, docs.
• Segmentation: Select line|paragraph|sentence|token and token chunk size for flexible ingestion.
• Pluggable retrievers: DeterministicRetriever (TF‑IDF/BM25 core) and a lightweight EmbeddingRetriever (Hashing + cosine).
• Evals framework scaffold: Pydantic models (Requirement, Proof, Assessment), runner, and assessment engine.
• UI/API: Gradio gets a “Regulatory Proof” tab and retriever/segmentation controls; API /reload now accepts segment/token_chunk_size/retriever.

Why it matters

• Reliability: Formal tests + fail‑closed retrieval build trust (especially for regulated use cases).
• Extensibility: Pluggable retriever and evidence model unlock future embedding/DB retrievers without changing the surface.
• Proof‑first evals: The proofcite.evals scaffold makes it straightforward to convert regulations into Requirements and produce audit‑ready Assessments.

Upgrade notes

• Install deps: pip install -r proofcite/requirements.txt
• Segmentation can lower per‑chunk scores; adjust --threshold accordingly (e.g., --segment sentence --threshold 0.2).
• Docker publish workflow will push on tags vX.Y.Z to DockerHub if DOCKERHUB_USERNAME and DOCKERHUB_TOKEN are set in repo secrets.

Try it

• Value report: python -m proofcite.examples.report_value
• Regulatory harness: python -m proofcite.regulatory --mode baseline --docs "proofcite/examples/regulatory/*.txt" --rules proofcite/examples/regulatory/rules_extended.jsonl --rerank hybrid --span_max_gap 1
• UI: proofcite-gradio (set PROOFCITE_DOCS or change glob in the field, pick retriever and segmentation).