Some HPE pages carry information only inside an image. The Morpheus
release schedule / support-lifecycle matrix (sf000111242en_us) is the
case that surfaced this: the entire "which version ships when, when a
stream hits Maintenance/EOL" table is a JPEG, so html_to_md dropped
every value and retrieval could never answer "when does 8.x.x reach
end of life".
New scrape/vision.py (opt-in via VISION_OCR) transcribes qualifying
images with a local Ollama vision model and appends a labeled markdown
block to the page so the text is chunked, embedded and retrieved.
Model/prompt chosen against ground truth on the matrix (14x4 cells):
- qwen2.5vl:7b + a bare "transcribe the table" prompt: 55-56/56, ~18s,
GPU-resident, and prompt-ROBUST (accurate without hand-tuning).
- gemma3:12b needs an exact "read row by row, keep cells aligned"
prompt or it shifts a column by one row; qwen2.5vl:32b is no more
accurate and 5x slower (CPU spill). Both documented in the code.
- Note: self-consistency (2 samples must agree) only catches RANDOM
flakiness — a wrong read is stable across seeds — so every block also
ships a "verify against the source image" caveat.
Reliability/cost:
- content-hash cache in corpus/.vision-cache/ (committed) — each unique
image OCR'd once ever, shared across version bundles.
- VISION_MAX_NEW bounds NEW OCRs per run so the first pass can't balloon
into hours; the cache fills incrementally. Deferred count is logged.
- every failure path degrades to the pre-vision behavior; never blocks
the scrape.
Also:
- add the morpheus_release_schedule bundle (sf single-doc) + 2 eval
golden queries for it.
- fetch_single_doc: title falls back to the bundle title (not docId)
when a page has no <h1> (sf solution articles).
- Pillow in requirements-vision.txt (scrape-only; kept out of the
server image), installed + VISION_* wired into refresh.yml.
Verified locally: matrix transcribes to the exact 14-row table; second
run is a cache hit (ocr=0).
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01LFowQzJu7k97QLCRDSAeh1
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) <[email protected]>
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) <[email protected]>