vecgrep

Semantic grep — like ripgrep, but with vector search.

Search your codebase by meaning, not just text. Ask for "error handling for network timeouts" and find the relevant code, even if it doesn't contain those exact words.

• Local-first — embedding model ships inside the binary. No API keys, no GPU, your code stays on your machine.
• Fast — indexes thousands of files in seconds, searches in milliseconds. Interactive TUI and HTTP server update results progressively during indexing.
• Bring your own model — optionally connect to Ollama, LM Studio, or any OpenAI-compatible API. See BENCHMARK.md.

Install

Pre-built binaries on the releases page, or:

mise use github:martintrojer/vecgrep  # mise
cargo install vecgrep                 # cargo

AI skill for coding agents

npx skills add martintrojer/vecgrep           # Codex / generic
/plugin marketplace add martintrojer/vecgrep  # Claude Code

Quick start

vecgrep "error handling" ./src                          # search by meaning
vecgrep --reindex --hybrid-index ./src                 # build lexical index support
vecgrep --hybrid "IndexConfig" ./src                   # optional lexical + semantic ranking
vecgrep -i "authentication"                             # interactive TUI
vecgrep -l "retry logic" | xargs rg "unwrap"            # combine with ripgrep
vecgrep --serve --port 8080 ./src                       # HTTP server for IDE plugins
vecgrep --embedder-url http://localhost:11434/v1/embeddings --embedder-model mxbai-embed-large "query"

Key features

Feature
TUI mode (-i)	Live search with preview pane, score colors, configurable file opener (--open-cmd)
HTTP server (--serve)	/search and /status endpoints for IDE integration (vecgrep.nvim)
Path scoping	Results scoped to requested paths, like ripgrep. --no-scope to search entire project
Hybrid search (--hybrid)	Optional lexical + semantic ranking for grep-like searches
Hybrid index (--hybrid-index)	Builds lexical index support required by hybrid queries
File type filters	-t rust, -T markdown, -g "*.rs" — same flags as ripgrep
JSONL output	--json for scripting with jq
Config files	Project (.vecgrep/config.toml) and global (~/.config/vecgrep/config.toml)
Cache	BLAKE3 content hashing — only changed files are re-embedded

Documentation

• User Guide — examples, path semantics, embedding models, ignoring files
• Configuration — config files, options reference, environment variables
• Server API — HTTP endpoints for IDE integration
• Benchmarks — model comparisons and retrieval accuracy
• Philosophy

How it works

1. Walk — .gitignore-aware file discovery (same engine as ripgrep)
2. Chunk — overlapping token-window chunks, snapped to line boundaries
3. Embed — built-in ONNX model or external API
4. Index — SQLite + sqlite-vec, keyed by content hash
5. Search — cosine similarity KNN, fast enough for every-keystroke use

License

MIT