zvec-ai/zvec-grep

English | 中文

Know the words—or don’t. Just zg.

The local-first search layer for humans and agents.

🎬 Tour | 💫 Features | 🚀 Try it yourself | 📚 Docs | 📊 Benchmarks | 🤝 Community

zg (zvec-grep), powered by zvec, unifies ripgrep, BM25, and vector search behind one local-first interface. Use it directly from the terminal, or let your agent use it for you.

This README and the documentation describe the TypeScript / Node.js implementation. The Rust implementation is developed separately in rust/.

• Ready for humans and agents — install once, index once, then use the same workspace from the CLI or your agent on macOS, Linux, and Windows.
• Search beyond keywords — discover by meaning, rank by relevance, then verify with exact text or regex when needed.
• Multi-format search — search source code, documents, and structured data while preserving useful structure and source locations.
• Less searching, less context — ranked, source-linked results surface the right evidence with fewer tool calls, fewer tokens, and less noise.
• Local by default — files, indexes, and local models stay on your machine; remote embeddings receive data only with your permission.

# Requires Node.js 22 or newer.
npm install -g @zvec/zvec-grep

mkdir zg-mystery && cd zg-mystery
curl --retry 3 --retry-all-errors --progress-bar -fL \
  -o alice-in-wonderland.txt https://raw.githubusercontent.com/GITenberg/Alice-s-Adventures-in-Wonderland_11/master/11.txt \
  -o sherlock-holmes.txt https://raw.githubusercontent.com/GITenberg/The-Memoirs-of-Sherlock-Holmes_834/master/834.txt

zg index --embedding local/potion-retrieval-32m

With OpenCode configured:

zg install --target opencode --yes
opencode models
opencode run --model opencode/nemotron-3-ultra-free \
  "An unseen creature left a few marks. What did the detective infer? Cite local evidence."

Free model availability can change. Check opencode models and replace the example model with one that is currently available in your environment.

OpenCode chooses zg on its own—the prompt does not name a tool.

⚙ zvec_grep_zvec_grep_search {"root":".../zg-mystery","query":"An unseen creature left a few marks. What did the detective infer?","fts":["marks","unseen creature","inferred","detective"],"fuse":true}
→ Read sherlock-holmes.txt [offset=5440, limit=120]

Answer:
The detective is Sherlock Holmes in *The Adventure of the Crooked Man*.
From the animal's foot-marks, he inferred:

- It was not a dog, cat, monkey, or other familiar creature: its print had
  five foot-pads, long nails, and was nearly as large as a dessert-spoon
  (sherlock-holmes.txt:5479-5486).
- It had a long body and very short legs: nearly two feet long, but with a
  stride of only about three inches (sherlock-holmes.txt:5497-5502).
- It was carnivorous because it climbed the curtain toward a canary's cage
  (sherlock-holmes.txt:5509-5510).
- It was probably from the weasel or stoat tribe, though unusually large
  (sherlock-holmes.txt:5514-5516).

Holmes also inferred that a man entered through the window with it and later
left with the key (sherlock-holmes.txt:5464-5470, 5527-5528).

Search the same bookshelf directly, without an agent:

zg query --human "An unseen creature left a few marks. What did the detective infer?" --limit 3

zg returns the relevant passages from sherlock-holmes.txt, ranked ahead of alice-in-wonderland.txt.

zg --index --index-embedding-concurrency <n> and ZVEC_GREP_INDEX_EMBEDDING_CONCURRENCY control embedding concurrency while building or updating an index, for both local and remote models. The environment variable also applies to automatic indexing and refresh. These controls do not change query-vector inference. The CLI option is accepted only with --index.

For llama.cpp, the limit controls contexts per indexing model instance. For Transformers.js, it controls calls in flight on one cached pipeline and does not guarantee simultaneous native/GPU execution. Both cap positive integer values at 8. For Potion/model2vec, it controls concurrent embedding batches without that cap; the default is 2 and the CPU worker pool has its own capacity limit. For remote models, both controls set the maximum concurrent batches without the cap of 8. Existing adaptive scheduling and its defaults remain unchanged; it may reduce concurrency after rate limits or retryable failures.

The priority is the explicit CLI/API index option, then the index environment variable, then ZVEC_GREP_LLAMA_CONTEXT_PARALLELISM (llama.cpp indexing only), then the automatic default.

For llama.cpp and Transformers.js, CPU execution or a runtime without a VRAM query uses 1 unless overridden. Transformers.js currently has no VRAM query, so its automatic limit is 1. When a GPU runtime provides free VRAM, the limit is floor(freeVRAM × 0.25 / 150 MiB), clamped to 1–8; a failed or invalid VRAM query uses 2. This retains the existing 150 MiB heuristic, which is not a guarantee that a model will fit in memory.

For CUDA errors, memory exhaustion, or native crashes, try a limit of 1 and retry the index. These examples run directly so the new environment takes effect immediately:

export ZVEC_GREP_INDEX_EMBEDDING_CONCURRENCY=1
zg --index --mode direct

Windows PowerShell:

$env:ZVEC_GREP_INDEX_EMBEDDING_CONCURRENCY = "1"
zg --index --mode direct

An explicit CLI option overrides the environment for that index operation, including when using the daemon; no daemon restart is needed for the CLI option:

export ZVEC_GREP_INDEX_EMBEDDING_CONCURRENCY=8
zg --index --index-embedding-concurrency 1

To change the daemon's environment-variable default, update its startup environment, then run zg --server off and zg --server on from that environment. If an agent launches zg, update its environment and restart the agent/MCP connection too. JavaScript exceptions can be caught, but native aborts can terminate the process before any CPU fallback runs. A limit of 1 reduces concurrency; it does not prevent every GPU failure. You can also retry with --device cpu.

Each benchmark uses paired A/B runs with tasks, agent/model, prompt, environment, and limits held constant; only zg access and usage guidance differ.

See the benchmark documentation for full results and reproduction details.

SWE-QA-Bench uses Claude Code with Claude Opus 5 at high reasoning effort; BrowseComp-Plus uses Codex gpt-5.6-sol at medium reasoning effort. Both zg profiles use Qwen3.7 Text Embedding.

• Why it helps: semantic discovery narrows the search space, ranked lexical retrieval anchors exact identifiers, and compact evidence reduces broad scans, repeated tool calls, and model context.
• Why it generalizes: the same retrieval loop works across domains—code is indexed with symbols, signatures, and breadcrumbs, while prose is retrieved as focused sections and chunks.

• Pylint — Python static analysis: the task asks how AST node handling separates annotated and non-annotated attribute initialization. Symbol-aware retrieval is useful because the architectural entry point is not known in advance.
• Matplotlib — plotting and rendering: the task traces FontInfo and font selection through multiple math-text rendering stages. Ranked semantic and lexical evidence helps reconstruct the cross-file data and control flow.
• Django — web framework: the task connects username uniqueness, ORM transactions, and formset bulk operations. Compact ranked evidence brings the distributed design rationale together.

zg works best when evidence spans files or modules and the target location is unknown, especially for call-chain, data-flow, and architectural questions. Since agents decide when and how to use it, results vary by model and run; repeated-run averages are more reliable.

💬 DingTalk	📱 WeChat	🎮 Discord	X (Twitter)
Scan to join	Scan to join	Click to join	Click to follow

Community contributions are always welcome—bug fixes, features, and documentation improvements all help make zvec-grep better.

Check out our Contributing Guide to get started!