kbolt

kbolt is a local-first retrieval engine for indexing local notes and docs and searching them with keyword, semantic, reranked, and deep retrieval modes.

Full documentation: https://h3nock.github.io/kbolt/

Install

macOS and Linux x86_64 with Homebrew:

brew install h3nock/kbolt/kbolt

Rust users on macOS, Linux, or Windows:

cargo install kbolt

Prebuilt binaries are also available from GitHub Releases.

If llama-server is not already installed, follow the official llama.cpp install guide.

Quick start

Set up the default local retrieval stack:

kbolt setup local
kbolt doctor

kbolt setup local downloads the default local embedder and reranker models, starts managed llama-server processes, and writes the local provider bindings into the kbolt config directory.

Add a folder of notes or docs:

kbolt collection add /path/to/docs --name my_docs

On macOS and Linux, keep it fresh automatically:

kbolt watch enable

Search the indexed content:

kbolt search "rust error handling"

Search modes

kbolt search "query" runs hybrid keyword + semantic retrieval. Reranking is opt-in for this default mode.

• kbolt search "query": hybrid keyword + semantic
• kbolt search "query" --rerank: hybrid + reranking (higher quality, slower)
• kbolt search "query" --keyword: keyword only
• kbolt search "query" --semantic: dense only
• kbolt search "query" --deep: explicit query expansion + multi-variant retrieval, reranked by default

kbolt setup local configures the default local embedder and reranker. To enable deep search later:

kbolt local enable deep

Use --deep when the query may not share vocabulary with the best matching documents, or when a short/underspecified query needs broader recall. It runs query expansion on every search and is slower than the default search and --rerank; for exact titles, named entities, or lexically clear lookups, start with normal search or --rerank.

What kbolt supports

• Index Markdown, plaintext, and source code (Rust, Python, JS/TS, Go, Java, Kotlin, C/C++, C#, Ruby, PHP, Swift) from one or more local directories
• Group collections into spaces and scope search with --space or --collection
• Search with keyword, semantic, hybrid reranked, and deep retrieval modes
• Read underlying source files with kbolt get, kbolt multi-get, and kbolt ls
• Check indexed content and disk usage with kbolt status
• Re-scan and re-index changed files with kbolt update
• Keep collections fresh automatically on macOS and Linux with kbolt watch enable
• Exclude files with gitignore-style patterns via kbolt ignore
• Check readiness with kbolt doctor and kbolt models list
• Run local models via llama-server or bind remote OpenAI-compatible endpoints through provider profiles
• Serve the index to agents over MCP with kbolt mcp
• Run retrieval benchmarks with kbolt eval ...
• Schedule recurring re-indexing with kbolt schedule ...