lilbee

Beta — feedback and bug reports welcome. Open an issue.

Local knowledge base that handles both code and documents, with a git-like per-project model. Search, ask questions, or chat — standalone or as a retrieval backend for AI agents via MCP. Everything stays on your machine, powered by Ollama and LanceDB.

• Why lilbee
• Demos
• Install
• Quick start
• Agent integration
• Interactive chat
• Supported formats
• Vision OCR (optional)
• Configuration
• How it works

Why lilbee

lilbee indexes documents and code into a searchable local knowledge base. Use it standalone — search, ask questions, chat — or plug it into AI coding agents as a retrieval backend via MCP.

Most tools like this only handle code. lilbee handles PDFs, Word docs, spreadsheets, images (OCR) — and code too, with AST-aware chunking.

• Standalone knowledge base — add documents, search, ask questions, or chat interactively with model switching and slash commands
• AI agent backend — MCP server and JSON CLI so coding agents (Claude Code, OpenCode, etc.) can search your indexed docs as context
• Per-project databases — lilbee init creates a .lilbee/ directory (like .git/) so each project gets its own isolated index
• Documents and code alike — PDFs, Office docs, spreadsheets, images, ebooks, and 150+ code languages via tree-sitter
• Open-source and fully offline — your documents never leave your machine. Runs with Ollama and LanceDB, no cloud APIs or Docker

Add files (lilbee add), then search or ask questions. Once indexed, search works without Ollama — agents use their own LLM to reason over the retrieved chunks.

Demos

Vision OCR

Standalone

Hardware requirements

lilbee runs entirely on your local machine — your hardware is the compute.

Ollama handles inference and uses Metal on macOS or CUDA on Linux/Windows. Without a GPU, models fall back to CPU — usable for embedding but slow for chat.

Install

Prerequisites

• Python 3.11+
• Ollama — the embedding model (nomic-embed-text) is auto-pulled on first sync. If no chat model is installed, lilbee prompts you to pick and download one.
• Optional (for image OCR): brew install tesseract / apt install tesseract-ocr

First-time download: If you're new to Ollama, expect the first run to take a while — models are large files that need to be downloaded once. For example, qwen3:8b is ~5 GB and the embedding model nomic-embed-text is ~274 MB. After the initial download, models are cached locally and load in seconds. You can check what you have installed with ollama list.

Install

pip install lilbee        # or: uv tool install lilbee

Development (run from source)

git clone https://github.com/tobocop2/lilbee && cd lilbee
uv sync
uv run lilbee

Quick start

# Check version
lilbee --version

# Initialize a per-project knowledge base (like git init)
lilbee init

# Chat with a local LLM (requires Ollama)
lilbee

# Add documents to your knowledge base (embedding runs locally — may take
# a moment per file, longer for large collections)
lilbee add ~/Documents/manual.pdf ~/notes/

# Ask questions — answers come from your documents via a local LLM
lilbee ask "What is the recommended oil change interval?"

# Search documents — returns raw chunks, no LLM needed at query time
lilbee search "oil change interval"

# Remove a document from the knowledge base
lilbee remove manual.pdf

# Use a different chat model
lilbee ask "Explain this" --model qwen3

# Check what's indexed
lilbee status

# Add a scanned PDF with vision OCR (prompts to pick a model if none configured)
lilbee add scan.pdf --vision

# Or set a vision model upfront and it applies to all future syncs
export LILBEE_VISION_MODEL=maternion/LightOnOCR-2
lilbee add scan.pdf

Agent integration

lilbee can serve as a local retrieval backend for AI coding agents via MCP or JSON CLI. See docs/agent-integration.md for setup and usage.

Interactive chat

Running lilbee or lilbee chat enters an interactive REPL with conversation history, streaming responses, and slash commands:

Slash commands and paths tab-complete. A spinner shows while waiting for the first token from the LLM.

Supported formats

Vision OCR (optional)

Scanned PDFs that produce no extractable text can be processed using a local vision model via Ollama. During add or sync, lilbee detects when text extraction yields nothing (common with scanned documents) and:

• Without a vision model configured: skips the file and tells you to re-add it with --vision or set LILBEE_VISION_MODEL
• With a vision model configured: rasterizes each page and sends it to the vision model for OCR. This is compute-intensive — expect seconds to tens of seconds per page depending on your hardware and model (see benchmarks below)

Setup:

# In chat, use the interactive picker:
/vision

# Or set directly:
/vision maternion/LightOnOCR-2

# Or via environment variable:
export LILBEE_VISION_MODEL=maternion/LightOnOCR-2

Recommended models:

Benchmarks: Apple M1 Pro, 32 GB RAM, Ollama 0.17.7. See benchmarks, test documents, and sample output.

Configuration

All settings are configurable via environment variables:

CLI also accepts --model / -m, --data-dir / -d, --vision-timeout, --log-level, and --version / -V flags.

How it works

Documents are hashed and synced automatically — add, change, or delete files and lilbee keeps the index current. Kreuzberg extracts text from PDFs, Office docs, images (OCR), etc. For scanned or image-heavy PDFs, lilbee can rasterize pages to images and run them through a local vision model via Ollama for higher-quality extraction. tree-sitter chunks code by AST. Chunks are embedded via Ollama and stored in LanceDB. Queries embed the question, find the closest chunks by vector similarity, and pass them as context to the LLM.

Data location

lilbee uses per-project databases when available, falling back to a global database:

1. --data-dir / LILBEE_DATA — explicit override (highest priority)
2. .lilbee/ — found by walking up from the current directory (like .git/)
3. Global — platform-default location (see below)

Run lilbee init to create a .lilbee/ directory in your project. It contains documents/, data/, and a .gitignore that excludes derived data. When active, all commands operate on the local database only.

License

MIT