Post-prefix eval (22 queries, 2026-09-30): dense MRR 0.966 vs
bm25+rerank 0.827. Watchtower keeps live RERANK_URL, so rerank
is gated on RERANK_ENABLED (default off) rather than URL presence.
mcp 2.0.0 removed `mcp.server.fastmcp`. This repo was pinned to <2
(6723829) to stop a rebuild from crash-looping the container the way
seed-mcp (2026-09-01) and zerto-docs (2026-08-11) did. This lifts the
pin by doing the port — the same mechanical change as
justin/seed-mcp#23 and justin/morpheus-docs#10.
- import: FastMCP -> MCPServer, plus TransportSecuritySettings
- constructor takes no transport options: MCPServer(f"{PRODUCT_NAME}-docs")
- host/port/stateless_http/transport_security are run() kwargs;
`mcp.settings` no longer exists
- requirement becomes plain `mcp>=2,<3` (2.x has no [fastmcp] extra)
@mcp.tool() decorators and all handler signatures are unchanged.
CI now smokes `import docs_mcp.server` twice in both workflows — once
after pip install, once inside the built image — so a green build can
never ship a non-importing container again.
Verified on python:3.12-slim (matches the image):
- tools/list dumped in wire format is BYTE-FOR-BYTE IDENTICAL between
mcp 1.30.0 and 2.2.0, all 10 tools
- streamable-http boots; initialize returns HTTP 200, both from
localhost and with a container-DNS `Host: hvm-docs-mcp:8000` header
(passing host= to run() keeps DNS-rebinding protection off)
- no mcp-session-id response header, i.e. stateless_http is in effect
- stdio boots; initialize + tools/list return 10 tools
- production image builds and imports under mcp 2.2.0; httpx 0.28.1 and
httpx2 2.12.0 coexist as expected
Retrieval is untouched, so no eval numbers are included.
Closes#12
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01QBx2P1VhKQEzJWoZ96QcnH
Phase 3/6/7/8 in one pass since they depend on each other.
* docs_mcp/server.py
- Wire search_docs / get_page / list_versions tool bodies.
- search_docs flow: BM25 first (rag.bm25 FTS5) → over-fetch RERANK_POOL
chunks → POST to RERANK_URL/v1/rerank → return top-k. Dense is the
fallback when BM25 finds nothing. HYBRID_SEARCH=true switches to
dense+BM25+RRF (fused via the new _rrf_fuse helper).
- All retrieval failures are caught and fall back to the next layer,
so a dead reranker or missing BM25 db never blocks a search.
- Source URLs built from the bundle's docId so results link straight
into support.hpe.com.
* eval/
- 22 hand-curated golden queries grounded in real corpus page titles.
- DenseRetriever / BM25Retriever / HybridRetriever / RerankedRetriever
+ MRR/Recall@K/nDCG@K harness. RERANK_URL env activates the
reranked variants.
- Committed eval/results/baseline.md. On this corpus:
dense: MRR 0.539
bm25: MRR 0.880
hybrid_rrf: MRR 0.692
bm25+rerank: MRR 0.920 (winner)
hybrid_rrf+rerank: MRR 0.875
HPE structured docs use controlled vocabulary, so lexical match
dominates. Hybrid loses because dense pollutes the fused pool.
* scripts/rerank_server.py
- Minimal HTTP /v1/rerank over sentence-transformers
cross-encoder/ms-marco-MiniLM-L-6-v2. Cohere-style request/response.
- This is the dev/CPU fallback; production replaces it with the
llama.cpp + jina-reranker-v2-base GGUF sidecar (same wire protocol).
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
claude
justinandClaude Opus 5