nicholasglazer/gnosis-mcp

[0.11.0] - 2026-04-18

Changed

• Default chunk size lowered 4000 → 2000 chars (GNOSIS_MCP_CHUNK_SIZE). Measured on a real 558-doc developer-docs corpus with 25 hand-written golden queries: 2000-char chunks sit on the peak nDCG@10 plateau (0.8702), up from 0.8416 at the old 4000-char default (+3 nDCG points, Hit@5 0.88 → 0.92). Existing corpora keep working; re-ingest with --wipe to regenerate at the new size. Full sweep in docs/bench-experiments-2026-04-18.md.
• Documentation hardened on reranker guidance. The bundled [reranking] cross-encoder remains off by default. New measurements on the same corpus: MiniLM rerank drops nDCG@10 by 27 points and adds 400× latency; BGE-reranker-v2-m3 drops nDCG@10 by 31 points and adds 2400× latency. Search skill now warns explicitly before enabling. Prose-trained rerankers fight reference-style docs; measure on your corpus before turning it on.

Added

• Release pipeline (scripts/bump-version.sh, scripts/release.sh, scripts/update-arch-sums.sh, scripts/render-llms.py, docs/releasing.md). Single-command version bump across 13 files — pyproject, __init__.py, server.json, marketplace.json, SECURITY.md, CHANGELOG.md, pkg/arch/PKGBUILD, pkg/arch/.SRCINFO, docs/rest-api.md, docs/show-hn.md, demo/hero.tape, skills/setup/SKILL.md, uv.lock. Extended scripts/check-versions.sh now gates all seven version-bearing files in CI and on every parity check. scripts/release.sh verify <X.Y.Z> polls PyPI / GHCR / MCP Registry / GitHub / AUR after a tag push and reports which are live.
• Post-release Arch-sums PR workflow. A new job in .github/workflows/publish.yml polls PyPI until the new sdist resolves, computes the tarball sha256, and opens a PR with the PKGBUILD + .SRCINFO update. First-time AUR publish walkthrough in docs/releasing.md §8.
• Template-rendered llms.txt / llms-full.txt. Source lives in llms.txt.tmpl and llms-full.txt.tmpl with {{VERSION}} / {{TEST_COUNT}} / {{MCP_MEAN_MS}} / {{MCP_P95_MS}} tokens; CI gates the rendered output for drift.
• Agents + skills refresh. agents/corpus-sync.md (bulk ingest / prune / wipe / crawl / git-history lifecycle with playbooks), updated agents/doc-keeper.md (single-file CRUD lane), and rewritten skills: skills/search/, skills/ingest/, skills/tune/, skills/manage/. Users can drop these straight into ~/.claude/agents/ and ~/.claude/skills/.
• Benchmark experiments log (docs/bench-experiments-2026-04-18.md): full chunk-size sweep (1000 → 4000 chars), reranker comparison table (keyword, MiniLM, BGE-v2-m3, mxbai-large), hybrid-vs-keyword on vocabulary-matched corpora, with raw numbers and methodology.
• tests/bench/bench_real_corpus.py --embed-model / --embed-dim flags — swap the embedder for A/B retrieval-quality shoot-outs against any model that ships an ONNX artefact on HuggingFace.
• GNOSIS_MCP_ALLOW_PRIVATE_CRAWL=true env override for the crawl SSRF guard — useful for local dev, CI, and Docker-internal testing. Mirrors the existing GNOSIS_MCP_WEBHOOK_ALLOW_PRIVATE pattern. Off by default.
• ONNX filename fallback in local_embed.py — tries onnx/model_quantized.onnx first, then onnx/model.onnx. Lets the embedder load any HF repo that ships either variant.
• /eval short-answer tooling exposed through the gnosis-mcp eval CLI and the tune skill — measures Hit@K / MRR / Precision@K on a user's own corpus in under a second.
• Docker image published to GHCR (ghcr.io/nicholasglazer/gnosis-mcp), multi-arch (linux/amd64, linux/arm64), built by .github/workflows/docker.yml on every tag push.
• Landing site refresh at gnosismcp.com: honest feature copy, fact-checked benchmark card, graph visualization driven by live documentation_links data, llms.txt + llms-full.txt served at the root for AI-assistant ingestion.
• Rewritten demo GIFs. demo/demo-hero.gif (440 KB) and demo/demo-crawl.gif (584 KB). Real commands only — no fake version echoes, no pip install during recording, no backgrounded-server noise. Source (VHS tape + staging scripts) lives in .internal/demo-source/.

Fixed

• llms.txt and llms-full.txt now reflect the actual test count (632) and the current SDK version.
• CHANGELOG ordering: [Unreleased] stays at the top; All notable changes… preamble moved out of the release block so it doesn't get rolled into a version section on the next bump.