POC: wire search page to semantic KNN search (sc-45806)

[nsantacruz]

Summary

• Disables Elasticsearch keyword search on the /search page (text search only) and instead renders results from the semantic KNN search pipeline.
• Up to 40 direct semantic matches, followed by up to 10 link-origin matches (refs frequently linked-to from the top semantic hits), each labeled with a chip in the result title (Semantic Match / Related Link).
• Filter sidebar and sort dropdown are hidden on semantic results — not useful for this few, non-facetable results.
• Sheet search is untouched (still keyword/Elasticsearch) since the KNN index only covers texts.

Why a new backend endpoint?

/api/knn-search is gated by a shared Authorization: Bearer <SEMANTIC_SEARCH_API_TOKEN> secret intended for server-to-server/tool callers. Calling it directly from browser JS would ship that secret to every visitor. Instead:

• KnnSearch.post() was split into a thin auth/dispatch wrapper and a reusable KnnSearch.run_search(body) classmethod (api/views.py).
• A new public, same-origin view semantic_search_wrapper_api (reader/views.py, routed at POST /api/search-wrapper/semantic) calls run_search directly server-side — no token required, same trust model as the existing search_wrapper_api.

Frontend

• New SemanticSearchQuerier.jsx (sibling to ElasticSearchQuerier.jsx) posts to the new endpoint, tags each hit resultOrigin: 'semantic'|'link', and feeds SearchPage.
• SearchResultList.jsx gets a new type === "semantic" render branch (avoids the !!result._source.version filter, which would silently drop every link-origin result).
• SearchTextResult.jsx renders the origin chip and falls back to a version-less href for link-origin rows (which lack version/lang).
• SearchPage.jsx hides the sort/filter controls when type === "semantic".
• ReaderPanel.jsx mounts SemanticSearchQuerier for text search, keeps ElasticSearchQuerier for sheet search.

Test plan

• webpack build succeeds across all bundles.
• Django imports cleanly; exercised /api/search-wrapper/semantic via Django test client — correctly calls Gemini to embed the query and reaches the pgvector query (failed only on local Postgres not running in this sandbox).
• Manual: docker-compose up, open /search?q=..., confirm no filter sidebar/sort dropdown, semantic results first with chips, link results below with chips, click-through navigates correctly.
• Manual: confirm sheet search on the search page is unaffected.

🤖 Generated with Claude Code

[gitvelocity-reviewer]

📊 Code Quality Score: 31/100

Base Score 51 × ESF 0.6 = 30.6 ≈ 31

Category	Score	Factors
🔭 Scope	13/20	9 files across backend (Python views, URL routing) and frontend (JSX, CSS); new public API endpoint; new React component; routing change affecting all non-sheet searches
🏗️ Architecture	12/20	Clean extraction of run_search() classmethod; new KnnSearchError exception type; SemanticSearchQuerier mirrors ElasticSearchQuerier pattern; bypasses existing auth pattern for same-origin calls which is an architectural risk
⚙️ Implementation	10/20	Refactoring of post() to classmethod is clean; result normalization is straightforward; conditional routing in ReaderPanel is simple but impactful; no complex algorithms, mostly plumbing and wiring
⚠️ Risk	12/20	All non-sheet text searches now route through semantic search with no feature flag; @csrf_exempt on public endpoint calling paid Gemini API with no rate limiting; deprecated React lifecycle method (componentWillReceiveProps); heRef bug indicates incomplete testing
✅ Quality	3/15	No tests added for new endpoint or classmethod refactor; POC comment in code suggests incomplete state; good docstrings on Python functions; heRef normalization bug indicates lack of testing
🔒 Perf / Security	1/5	No rate limiting on new public endpoint; Gemini API cost exposure via unauthenticated endpoint; no monitoring or alerting added

Was this score accurate? 👍 Yes · 👎 No

How this was scored →

_{Scored by GitVelocity · How are scores calculated?}