paparats-search

Searching code with Paparats

Paparats indexes the user's repositories into chunks (functions, classes, blocks) with a symbol graph between them. Its tools find code by meaning, not by exact text.

Which tool

• "Where is the code that does X?" / "How does X work?" — search_code with a plain-language question. Then get_chunk on the best hits instead of reading whole files.
• "Who calls this?" / "What uses this?" — find_usages on a chunk from a search result. It follows real symbol edges, not name matches.
• An exact identifier, error string or file name you already know — a text search (git grep -n, rg) is usually faster and complete.

Results are a ranking, not a complete list

search_code returns the best matches, capped by limit. It never proves that something does not exist, and it does not list every occurrence. For questions that need a complete answer — "every place that reads this env var", "all callers of X" — use the search to learn the right names, then confirm the full set with find_usages or a text search.

Reading the scores

Each result carries a relevance score. Treat 60% and above as a confident match; 40–59% as a lead to verify; below 40% as a sign the index has nothing close — rephrase once, then fall back to a text search.

Recent edits may not be indexed yet

The index is refreshed on a schedule (by default the indexer checks for changes about every 10 minutes), not on every save. For code the user has just written or changed, read the file or use a text search instead of relying on search_code.

When the tools are missing

The tools come from a Paparats server running on the user's machine (default http://localhost:9876/mcp). If they are unavailable, tell the user to check the stack with paparats status rather than guessing at the code.