agent-library 0.7.0

Markdown document management system with vector and full-text search

Librarian

A personal knowledge library for AI agents, built on Arcade for the Model Context Protocol (MCP).

Overview

Librarian provides AI agents with persistent storage for text, documents, and knowledge. Agents can store information and retrieve it later through semantic and keyword search, maintaining context across conversations.

graph LR
    A[Agent Stores Info] --> B[Parser]
    B --> C[Chunker]
    C --> D[Embedder]
    D --> E[(SQLite + vec)]
    F[Agent Queries] --> G[Hybrid Search]
    E --> G
    G --> H[Relevant Context]

Features

• Persistent knowledge storage for AI agents
• SQLite storage with sqlite-vec for vector search
• Full-text search using FTS5 with BM25 ranking
• Hybrid search combining semantic and keyword matching
• Max Marginal Relevance (MMR) for diverse results
• Configurable embedding models (local or OpenAI-compatible API)
• Header-aware text chunking with overlap
• Time-bounded search filters
• CLI and MCP server interfaces

Installation

git clone https://github.com/ArcadeAI/librarian.git
cd librarian
./setup.sh

Or install manually:

uv pip install -e ".[dev]"

CLI Usage

# Add files to the library
libr add ~/notes

# Search the library
libr search "machine learning concepts"

# List sources
libr list

# View library statistics
libr index

# Rebuild the index
libr index build

MCP Server

Start the server for AI assistant integration:

# stdio transport (Claude Desktop, CLI)
libr serve stdio

# HTTP transport (Cursor, VS Code)
libr serve http --port 8000

See the Arcade MCP documentation for integration details.

Available Tools

Tool	Description
Librarian_SearchLibrary	Search the library with hybrid vector + keyword search
Librarian_SemanticSearchLibrary	Find content by meaning (semantic similarity)
Librarian_KeywordSearchLibrary	Find content by exact keywords
Librarian_SearchLibraryByDates	Search within a date range
Librarian_AddToLibrary	Store new content in the library
Librarian_UpdateLibraryDoc	Update existing content
Librarian_ReadFromLibrary	Read full document content
Librarian_RemoveFromLibrary	Remove content from the library
Librarian_ListLibraryContents	List all stored content
Librarian_IndexDirectoryToLibrary	Bulk import files
Librarian_GetLibrarySources	List sources with document/chunk counts
Librarian_GetLibraryStats	Overall library statistics

Configuration

Set via environment variables:

Variable	Default	Description
DOCUMENTS_PATH	./documents	Root directory for files
DATABASE_PATH	~/.librarian/index.db	SQLite database location
EMBEDDING_PROVIDER	openai	local or openai
EMBEDDING_MODEL	all-MiniLM-L6-v2	Local model name
OPENAI_API_BASE	http://localhost:7171/v1	OpenAI-compatible API URL
OPENAI_EMBEDDING_MODEL	qwen3-embedding-06b	API model name
CHUNK_SIZE	512	Max characters per chunk
CHUNK_OVERLAP	50	Overlap between chunks
SEARCH_LIMIT	10	Default results limit
MMR_LAMBDA	0.5	MMR diversity (0=diverse, 1=relevant)
HYBRID_ALPHA	0.7	Vector vs keyword weight (1=vector only)

Project Structure

librarian/
├── cli.py           # Command-line interface
├── server.py        # MCP server and tool definitions
├── config.py        # Configuration management
├── indexing.py      # Document indexing service
├── types.py         # Shared type definitions
├── storage/
│   ├── database.py  # SQLite operations
│   ├── vector_store.py  # sqlite-vec search
│   └── fts_store.py     # FTS5 search
├── processing/
│   ├── embed/       # Embedding providers
│   ├── parsers/     # Document parsers
│   └── transform/   # Text chunking
├── retrieval/
│   └── search.py    # Hybrid search + MMR
└── utils/
    └── timeframe.py # Time filter utilities

Development

make install    # Install dependencies
make test       # Run tests
make lint       # Run linter
make format     # Format code
make typecheck  # Type checking
make check      # All checks
make evals      # Run evaluations

Resources

• Arcade.dev - Build AI-native applications
• Arcade Documentation - Integration guides and API reference

License

MIT License - see LICENSE for details.

Contact

• Email: [e-mail]
• Website: arcade.dev