22 queries against the prod image index rebuilt today on the expanded
GPU pool with the resilient embedder (PR #8): dense MRR 0.539→0.557,
bm25+rerank 0.920→0.959, hybrid_rrf+rerank 0.875→0.960 vs the
2026-05-22 baseline. No regression from mixed-provenance embeddings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Match the OLLAMA_URLS pattern from zerto-docs-rag so every docs MCP
build fans out across the same two GPU-pinned Ollama containers on
192.168.0.2 (:11435 Titan X, :11436 1080 Ti). The host's primary
Ollama on :11434 is left alone for OpenWebUI.
rag.embeddings now reads OLLAMA_URLS (plural CSV) preferentially with
fallback to OLLAMA_URL, defaulting to http://192.168.0.2:11434 — same
shape as zerto's embeddings.py. The OllamaEmbeddings class already
round-robins per batch, so both GPUs run in parallel during the
chroma rebuild.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 4/5 — adapt the template workflows to Justin's self-hosted Gitea
+ act_runner setup (see reference_gitea_server memory):
* PUSH via LAN endpoint 192.168.0.2:1234 (bypasses Cloudflare's 100 MB
request-body cap on the Free plan); PULL via git.jpaul.io.
* buildx with config-inline insecure-registry for the LAN endpoint —
docker/login-action can't be used there (host daemon rejects HTTP).
Auth is written into ~/.docker/config.json so buildx reads it
directly.
* docker/metadata-action labels org.opencontainers.image.source with
the PUBLIC URL so Gitea auto-links the package; explicit POST to
/api/v1/packages/.../-/link/{repo} as belt-and-suspenders (201 newly
linked, 400 already linked, both treated as success).
* deploy/docker-compose.yml: substitute <product> placeholders, point
image at git.jpaul.io/justin/hvm-docs:latest, set HYBRID_SEARCH=false
to match the eval winner (bm25+rerank), keep the llama.cpp + jina
GGUF reranker sidecar as the production target.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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) <noreply@anthropic.com>
The chunker emits any single paragraph as a stand-alone chunk regardless
of size. One HVM page had a 14,858-char paragraph (a big config table) —
nomic-embed-text 400'd the entire embed batch because the model's context
is 2048 tokens. Added a hard-split fallback that splits any oversized
chunk on line boundaries to MAX_CHARS=6000 (~1500 tokens, headroom).
Also defaulted PRODUCT_NAME to "hvm" in rag/index.py to match server.py.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 1: scrape User Manual (8.1.0/.1/.2), Release Notes (8.1.0/.1/.2),
and the unversioned Deployment Guide. Total ~1,160 pages, 9.7 MB markdown.
Discovers via the anonymous JSON API at /hpesc/public/api/document/{docId}:
/toc walks the page tree (for TOC-paginated docs), /render?page=GUID
fetches per-page HTML, /document/{docId} returns the whole body for
single-doc shapes like Release Notes.
Runner converts DITA-source HTML to clean markdown (strips Notices/
Acknowledgments/Abstract boilerplate), writes corpus/<bundle>/<page>.{md,json},
then a finalize pass synthesizes topic_cluster.clustered_topics by GUID
overlap across versions (HPE GUIDs are stable cross-version — confirmed
374/376/376 with 100% overlap on shared pages).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First dispatch on the empty template failed at Chroma collection
creation because PRODUCT_NAME was the literal string "<product>"
(YAML doesn't expand placeholders), and Chroma rejects collection
names containing characters outside [a-zA-Z0-9._-]:
chromadb.errors.InvalidArgumentError: Validation error: name:
Expected a name containing 3-512 characters from [a-zA-Z0-9._-],
starting and ending with a character in [a-zA-Z0-9]. Got:
<product>_docs
Same fix as the IMAGE env: derive from the repo name dynamically
via ${{ github.event.repository.name }}. Cloners can still override
explicitly, but a fresh clone now runs the index-rebuild step
cleanly out of the box.
Verified by re-dispatch — should fail next at docker login (placeholder
REGISTRY_PUSH hostname), which is the next-expected fail point and a
real per-deployment config the cloner has to fill in.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Both workflows had a static IMAGE env (<owner>/<product>-docs-mcp)
and a static --package arg in the GC step. Switch both to Gitea
Actions context variables so a clone of the template into any repo
name works on the first CI run without find/replace:
IMAGE: ${{ github.repository_owner }}/${{ github.event.repository.name }}
--owner ${{ github.repository_owner }}
--package ${{ github.event.repository.name }}
Also add the "Link container package to this repo" step that was
missing from the template (and which, naively copy-pasted from the
reference build, would have linked everything back to docs-mcp-
template). The new step derives owner + package + link-target all
from the running repo's context.
The github.* namespace is Gitea Actions' inherited GitHub-Actions
context — values come from the Gitea server, not github.com. Same
mechanism the reference build's $GITHUB_SHA tag-builder uses.
CLAUDE.md updated to note that image and package naming are
repo-derived; only registry endpoints and the Ollama URL need
per-clone editing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
justin
hvm-docs-refresh