Skip to content

Optional LanceDB Search Lifts Reported Recall From 58.2% to 71.4%

Sep 11, 2026
GitHub
Article image for Optional LanceDB Search Lifts Reported Recall From 58.2% to 71.4%

Summary

A local code-analysis system now pairs token search and graph expansion with optional LanceDB vectors, lifting reported recall from 58.2% to 71.4% while its SHA-256 ingest queue skips unchanged files and retries failures up to three times.

Key Points

  • The two-step ingest pipeline hashes source files with SHA-256, skips unchanged inputs, persists a serial queue across restarts, and retries failed tasks up to three times.
  • Optional LanceDB vector search raises reported overall recall from 58.2% to 71.4%, while retrieval combines token search, graph expansion, and configurable 4K-to-1M-token context budgets.
  • A token-protected local API on 127.0.0.1:19828 exposes project, file, review, hybrid-search, graph, rescan, embedding, and Agent-chat endpoints; a bundled MCP server lets Claude Code and Codex use them.

Tags

Read Original Article