fix(knowledge): optimize search and trace pipeline latency

[icecrasher321]

Summary

• Keep broad permission-filtered vector searches on HNSW in the shared query layer. Probe visible IDs without reading vectors, rank small scopes by those bounded IDs, and restart exact fallback from its own cursor for underfilled approximate pages. Deduplicate previously considered candidates and preserve distance ranking across refills. Live source authorization still gates content.
• Add correlated pipeline timings for Assistant and Search, including active-stage reports for stalled calls, provenance, billing, provider authorization, SQL, and result presentation. Record UTF-8 output/passage byte counts without content, and correct the final tool trace to measure bytes rather than characters.
• Add an opt-in PostgreSQL/pgvector harness that calls the real Assistant tool and compares identical Search-tab requests across a 30,000-chunk synthetic corpus, selective permissions, revocation, concurrent calls, and organization scope. Correct the ACL-transition test to cover pending and completed rewrites for candidate and final-read predicates.

Type of Change

• Bug fix

Testing

• Synthetic paired SQL benchmark: median vector time including the new probe fell from 81.5 ms to 3.8 ms; buffer accesses fell from 123,278 to 2,403. All 50 returned candidates matched exact ranking in this corpus. These are local SQL measurements, not end-to-end or deployment latency claims.
• 8 real PostgreSQL/pgvector harness scenarios passed again after review fixes, including saved pipeline diagnostics, passage-size assertions, and matching Search-tab/Assistant SQL. The synthetic 15-passage response was about 54 KB (45 KB of text); local presentation and measurement took 0.4–1.7 ms. No model latency claim is made from this provider-controlled harness.
• 198 search, tool-execution, and diagnostic tests passed for the review fixes, including exact-cursor restarts, deduplication, cross-page ranking, and UTF-8 size reporting.
• 224 search/application/diagnostic unit tests, 20 database mock tests, and 66 broader access integration tests passed; one existing integration test was skipped.
• Full lint, all 26 workspace type checks, all 46 repository audits (including API validation), block-registry validation, generated artifacts, and docs-manifest checks passed.

Run the harness from the repository root with Docker running:

KNOWLEDGE_SEARCH_PERFORMANCE_REPORT_FILE=/tmp/search-performance.json bun run test:search-performance

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
docs	Ready	Preview	Sep 11, 2026 8:06pm UTC

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; no actionable new correctness, security, or repository-rule violation remains.

Summary

• Keeps broad searches on HNSW while ranking bounded permission scopes exactly.
• Gives approximate and exact ranking strategies independent cursors, deduplicates considered candidates, and preserves distance ordering across refills.
• Adds stage timing and UTF-8 byte-count telemetry without logging queries or passage content.
• Adds an opt-in PostgreSQL/pgvector performance and authorization harness plus targeted regression coverage.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Search request] --> B[Scope and authorization]
  B --> C[Embedding]
  C --> D[Permission-filtered candidate probe]
  D -->|Broad scope| E[HNSW ranking]
  D -->|Bounded scope| F[Exact ranking by candidate IDs]
  E --> G[Live source authorization and hydration]
  F --> G
  G -->|Underfilled ANN page| H[Exact fallback with independent cursor]
  H --> G
  G --> I[Deduplicate and distance-rank]
  I --> J[Present results]
  J --> K[Content-free timing and byte diagnostics]

Loading

_{Reviews (2) · Last reviewed commit: "fix(knowledge): preserve fallback rankin..."}

[icecrasher321]

@greptile

[icecrasher321]

@cubic-dev-ai review this PR

[cubic-dev-ai]

@cubic-dev-ai review this PR

@icecrasher321 I have started the AI code review. It will take a few minutes to complete.