Search a codebase

A coding assistant has a hard limit on what it can hold in context, and it blows past that limit on any real codebase. Point nidus at the repo instead: it chunks by function, struct and trait rather than by character count, and every hit comes back as a symbol with the lines it lives on, never the body itself.

What you need

• Code search: AST-aware chunking, one chunk per symbol, results grouped by file.
• MCP: one line connects the store to Claude Code, Cursor, or anything else that speaks the protocol, with no glue code.
• Full-text search (BM25): the run below never called an embedding provider; BM25 keyword matching over the chunked text is what answered it.

Doing it

read the repo, then ask it something

nidus code ingest . --dir ./store

nidus code search "where do we fsync the write ahead log" \
  --dir ./store

[
  { "path": "store/write.rs", "language": "rust",
    "symbols": [
      { "symbol": "commit", "kind": "function",
        "start_line": 136, "end_line": 148, "score": 10.09 },
      { "symbol": "maybe_sync", "kind": "function",
        "start_line": 109, "end_line": 119, "score": 7.75 }
    ] }
]

138 files became 3,403 symbols in 1.7 seconds, with no API key, and it found the two functions that do the work.

hand it to your assistant

claude mcp add nidus -- nidus mcp --dir ~/.nidus \
  --embed-provider voyage --embed-model voyage-4

That is the entire integration. No SDK to install, no glue code to write.

What to tune

• Adding an embedder: search by meaning instead of exact keywords once a provider is configured.
• Summarize-then-embed: embed what a symbol means rather than its literal tokens, for a query that never spells the words the code itself uses.

Where it stops

Results are symbols and line spans, never source bodies: the assistant always opens the file for the real text. code is part of the default build; --no-default-features excludes it along with the rest of the ingest layer.