rag-jev 0.2.0 — inspectable RAG context selection

rag-jev adds inspectable Jev filtering and reranking between an existing retriever and answer model. This first public release is MIT licensed and includes Python, HTTP, TypeScript, LangChain and Dify integrations.

• Key-free workbench with recorded improvement and regression examples, policy replay, and import/export.
• Independent/contextual scoring, shadow mode, pins, evidence groups, and passage token budgets with explicit pinned-budget overruns.
• Calibration CLI that freezes a development-selected policy before held-out evaluation; reusable shadow configuration and family-cluster uncertainty estimates.
• Configurable OpenAI-compatible answer comparison, sanitized errors, bounded concurrency, and private vulnerability reporting.
• Blinded human-review interface and label validation. Independent review remains incomplete: 0/80 answers.

Validation: 93 Python tests and 6 TypeScript tests passed locally and in GitHub CI on Python 3.11 and 3.13. Clean wheel/npm installations passed. The frozen robustness archive reproduced byte-identical results in a clean environment. Its 108 live synthetic selection calls had no errors/bypasses; these are diagnostics, not production latency or security guarantees.

Research: on 300 BEIR/SciFact queries, observed NDCG@10 was 75.13 Jev / 72.11 Ettin / 66.47 BM25 from identical BM25 top20 candidates. On isolated HotpotQA/MuSiQue answer evaluations, normalized estimated API costs fell 19.7%/22.8%, including Jev. Both answer-F1 difference intervals include zero: the accuracy-superiority gate did not pass. Benchmark overlap and other limits are detailed in the repository reports. These results do not establish universal accuracy or total-cost gains.

Install the release packages

Published packages: PyPI and npm. Install from the registries:

python -m pip install rag-jev==0.2.0
rag-jev serve --replay-only
# Optional TypeScript client:
npm install rag-jev-client@0.2.0

Open http://127.0.0.1:8000. Live scoring requires a TypeSafe key and sends query/passage text to its API. Both registry packages passed clean installation checks. The npm install passed all six client tests, and the PyPI install passed the replay-only service smoke. PyPI publication used GitHub trusted publishing; npm used an authenticated, security-key-approved CLI publish. Downloadable wheel and npm tarball are also attached below.

Research artifacts

• reproduction.tar.gz: immutable round-3 answer and SciFact study snapshot, raw records and verification evidence. Use its included code for historical reproduction.
• robustness-reproduction.tar.gz: separate frozen synthetic workload and raw selection records.
• SHA256SUMS: checksums for release attachments; verify with shasum -a 256 -c SHA256SUMS after downloading all attachments.
• Code is MIT; benchmark material retains upstream terms described in THIRD_PARTY_NOTICES.md.

The archives contain separate historical source snapshots. Do not apply their frozen fingerprints to the current product checkout.