Codebase RAG

Semantic search over whatever project you're currently working in. The server detects the project root, keeps an index in a per-machine cache, and watches the filesystem so the index stays current as files change.

When to use it

• Before editing unfamiliar code: search for related implementations.
• When looking for callers, callees, or similar patterns.
• When the user's instruction doesn't say which file to touch.
• Before modifying a file: check what depends on it.

Tools

rag_search(query, num_results=5, source_type="all")

Semantic search across constraints, patterns, and code. Returns ranked results with file paths and relevance scores.

source_type is one of "all" (default), "docs" (constraints + patterns), "code" (source only), or "constraints" (rules only).

rag_query_impact(file_path, num_similar=5)

Show what depends on a file: its exports, the files that import it, and semantically similar files that may need coordinated changes. Pass the path relative to the project root, with forward slashes.

Choosing source_type

The default "all" is correct most of the time. Narrow it when:

• "constraints" — checking what rules govern an area before designing a change. Returns only architecture / constraints / CLAUDE.md-style files.
• "docs" — understanding intent and patterns before reading code. Returns constraints plus docs/patterns/ style documentation.
• "code" — finding existing implementations to copy or call. Filters out docs so examples aren't drowned out by prose.

A useful two-step for non-trivial changes: search "docs" first to learn the rules, then "code" to find concrete implementations.

Workflow

1. rag_search("what you're about to do") — read the top hits.
2. If you're going to modify a specific file, rag_query_impact("path/to/file").
3. Make the change.

How retrieval is weighted

The server gives higher retrieval weight to files it treats as constraints or patterns, so they outrank random source matches:

• ARCHITECTURE.yml / ARCHITECTURE.yaml — 10×
• CONSTRAINTS.md — 10×
• CLAUDE.md — 10×
• Anything under docs/patterns/ — 8×
• Everything else — 1×

If the project has rules or patterns worth surfacing first, putting them in one of these files (or under docs/patterns/) is the simplest way to make them appear at the top of search results. No config needed.

Adding custom sources

To index ADRs, OpenAPI specs, style guides, RFCs, or any other docs the project already maintains, edit customSources in the per-project config.json (see "Where state lives" below):

{
  "customSources": [
    { "pattern": "docs/adr/*.md",        "sourceType": "docs",        "weight": 9.0 },
    { "pattern": "openapi.yaml",         "sourceType": "constraints", "weight": 10.0 },
    { "pattern": "docs/guides/**/*.md",  "sourceType": "docs",        "weight": 8.0 }
  ]
}

• pattern — glob relative to the project root.
• sourceType — "constraints", "docs", or "code". Determines which collection the files index into and which source_type filter surfaces them.
• weight — retrieval weight. Built-in constraints use 10.0, patterns use 8.0, code uses 1.0.

The next rag_search call rebuilds the index to pick up the change.

Where state lives

The index and config live in a per-machine cache directory keyed by project root, not inside the project tree:

• Windows: %LOCALAPPDATA%\codebase-rag\<hash>\
• macOS: ~/Library/Caches/codebase-rag/<hash>/
• Linux: $XDG_CACHE_HOME/codebase-rag/<hash>/ or ~/.cache/codebase-rag/<hash>/

Inside that directory: config.json (project config + custom sources) and collections/ (ChromaDB on-disk store). Deleting the directory forces a fresh index on next query.

Status responses

Both tools return a status field. Handle the non-success cases:

• "success" — results are valid.
• "indexing" — first-run index is still building. Wait a few seconds and retry the same call. Result lists are empty until ready.
• "no_project" — no project root detected (no .git, package.json, pyproject.toml, Cargo.toml, or go.mod walking up from cwd). Tell the user; the server can't operate without a project root.
• "index_failed" — bootstrap failed. The summary field has the reason. Surface it; don't retry blindly.

First call in a brand-new project

The very first query in a project that's never been indexed takes a few seconds while the index builds (you'll see status: "indexing"). After that, queries are sub-second and the index stays fresh on its own via the filesystem watcher.