Offline Semantic Search Over Your Own Code

Ask an AI assistant a question about your codebase and watch what it does first: it starts reading. It greps for a term, opens a handful of files, follows an import, opens more. Most of the token bill — and most of the wait — is the assistant re-deriving structure that could have been handed to it. And plain text search has a second problem: it only finds names you already know. "Where do we verify webhook signatures?" doesn't match if the function is called checkHmac.

Vibgrate AI Context fixes both. It builds a local semantic index of your code once, then serves it to your assistant so every question is answered by querying meaning — offline, with no API key and nothing uploaded.

Grep finds names; you often don't know the name

Lexical search is exact and fast, and that is also its ceiling. It answers "where is checkHmac" but not "where do we verify webhook signatures," because it has no notion that those describe the same thing. Semantic search closes that gap: it ranks code by what it does, so a plain-English question lands on the right function even when your vocabulary and the code's vocabulary differ.

vg embed: a local semantic index

Build the index once. vg embed computes embeddings for the nodes in your code map using a model that runs entirely on your machine — no service call, no key:

$ vg embed
✔ vg embed · semantic index ready (1,204 nodes · local model)

The model is cached centrally under your home cache directory, so a second project reuses it instead of downloading again, and the per-repository vectors live in .vibgrate/cache. If you want the hard guarantee that nothing leaves the machine, run any command with --local; semantic embedding is skipped and search falls back to lexical only.

vg ask: query by meaning

With the index built, ask in plain English. vg ask blends lexical, structural, and semantic retrieval and returns a compact, ranked answer with a token budget you control:

$ vg ask "where do we verify webhook signatures?"
vg · 3 matches · ~2,000 tokens (budget 2000) · semantic
  match  src/webhooks/verify.ts:42  verifySignature()

The default budget keeps answers small enough to paste into a prompt without drowning the model. Pass -b to change it, or --no-semantic to force a purely lexical pass.

vg serve: hand it to your assistant over MCP

You don't have to run vg ask by hand. vg serve starts Vibgrate AI Context — a local-first MCP server that exposes your code map, drift, and version-correct library docs to any assistant that speaks the Model Context Protocol: Claude, Codex, and other MCP clients.

$ vg serve
vg · MCP server on stdio (read-only, auto-refresh).
     Connect your assistant to this process.

Over MCP the assistant gets structured tools instead of a shell: a symbol-and-text lookup for when it knows the name, a meaning-based query for when it doesn't, and graph navigation to trace callers and callees. It reads the map you already built rather than re-reading your files — which is exactly the mechanism behind our measured ~24% cut in AI coding tokens. Run it with --http and --port if your client connects over HTTP, or under --local for a full air-gap.

vg lib add: version-correct docs, ingested locally

Assistants also reach for the wrong documentation — the latest docs for a library you pinned two minors ago. Vibgrate AI Context can serve docs that match the version in your lockfile, and you can add your own. vg lib add ingests documentation from a local path, a URL, or a git repository:

$ vg lib add ./docs
$ vg lib add https://github.com/acme/widgets --online

This is the same problem we covered in why your AI assistant writes deprecated code — the fix is better context, not a bigger model.

What stays on your machine

Three plain facts kill three objections. The embedding model runs locally, so there is no API key and no per-query cost. The index and vectors are written under .vibgrate in your repo, so nothing is uploaded. And under --local, Vibgrate AI Context makes no network calls at all — suitable for an air-gapped or regulated environment.

Try it

$ vg embed
$ vg serve

Then point your assistant at the server. The Vibgrate AI Context reference covers every flag, and the free developer tools page shows what you get without an account.