Gemdex code search

Gemdex code search

This repository is indexed by Gemdex, a semantic code search MCP server (Gemini embeddings + Milvus). Reach for it first for anything that's a question about meaning, intent, or behavior — and only fall back to Grep/Glob for exact-string lookups.

When to use Gemdex vs Grep / Glob

Prefer the search_code MCP tool (provided by the gemdex server) when:

• The user asks about code by intent or meaning — e.g. "find the retry-with-backoff helper", "where does JWT refresh happen?", "websocket reconnection logic", "rate limiting".
• You want ranked, semantically-related results across many files.
• You're exploring an unfamiliar codebase and don't yet know the symbol or file names.

Use Grep / Glob only when:

• You need an exact string match (a specific symbol you already know, an error literal, a log message, an import path).
• You need to list files by path pattern.
• Gemdex returned no useful results for the semantic query.

Workflow

1. First time in a fresh checkout? Call index_codebase once with the repo's absolute path. After that, the index stays in sync automatically — this plugin's PostToolUse hook touches ~/.gemdex/.sync-trigger after every Edit, Write, or MultiEdit, which the gemdex server picks up and uses to incrementally re-embed only what changed.
2. For semantic questions, call search_code with a natural-language query and the absolute codebase path. Read the returned file:line hits first instead of guessing or grepping.
3. If results look stale or empty, call get_indexing_status to check the last sync time and chunk count. If the codebase isn't covered, run index_codebase again.
4. Fall back to Grep/Glob only for exact-string lookups (function names you already have, log lines, error strings, file globs).

Tool reference

The gemdex MCP server exposes four tools:

• search_code(path, query, limit?) — semantic search; returns ranked file:line snippets. Default limit is 5; raise to 10–15 when scoping out an unfamiliar area.
• index_codebase(path) — initial full index of a directory.
• get_indexing_status(path) — last sync time, chunk count, and freshness for a codebase.
• clear_index(path) — drop the index for a codebase.

Always pass an absolute path for path.