Sirchmunk offers an "embedding-free," real-time intelligence pipeline that bypasses traditional RAG system costs and rigidity. It processes raw data directly, providing agile insights for AI agents and power users, featuring a self-evolving knowledge base that adapts to data changes.

How It Works

This system eschews vector embeddings for direct, instant, full-fidelity raw data search. Its core is "Monte Carlo Evidence Sampling," a token-efficient, three-phase strategy for intelligent document region sampling and LLM processing. Search outputs form "Self-Evolving Knowledge Clusters" that dynamically update with queries. This query-driven evolution enables semantic broadening and near-instant retrieval of cached information, eliminating re-indexing and LLM calls for repeated queries.

Quick Start & Requirements

Install via pip install sirchmunk or uv pip install sirchmunk. For the web UI, use pip install "sirchmunk[web]". Prerequisites: Python 3.10+, an OpenAI-compatible LLM API key, and optionally Node.js 18+ for UI builds. ripgrep-all and ripgrep are auto-installed or require manual setup. Documentation is at https://modelscope.github.io/sirchmunk-web/.

Highlighted Details

• Embedding-Free Retrieval: Direct search on raw files, bypassing vectorization overhead for immediate, lossless data access.
• Self-Evolving Knowledge Clusters: Dynamically updated knowledge units learn from queries, enabling query-driven embeddings and accelerating similar future searches.
• Monte Carlo Evidence Sampling: Novel, token-efficient method for deep, context-aware evidence extraction.
• MCP Integration: Exposes search as MCP tools for seamless integration with AI assistants (e.g., Claude Desktop, Cursor IDE).
• Multi-Mode Search: Supports FAST (default, rapid), DEEP (comprehensive analysis), and FILENAME_ONLY (metadata) modes.

Maintenance & Community

Active development is evident from early 2026 releases. Hosted by modelscope on GitHub, community interaction primarily occurs via GitHub issues and discussions.

Licensing & Compatibility

Licensed under Apache License 2.0, permitting commercial use and integration into closed-source projects.

Limitations & Caveats

Current limitations include missing web search integration, multi-modal support, and distributed search, as per the roadmap. Core search functionality (except FILENAME_ONLY) requires an LLM API key, and the project appears to be in an early development stage.