modelscope/sirchmunk

What it solves

Sirchmunk addresses the rigidity and high overhead of traditional vector-based retrieval (RAG). It eliminates the need for expensive pre-processing, static vector embeddings, and complex database infrastructure, which often lead to stale data and information loss during approximation.

How it works

Sirchmunk uses an "indexless" approach that works directly with raw data. It employs agentic search and budgeted evidence exploration to retrieve information in real-time. The system features a self-evolving knowledge engine that automatically updates its knowledge graph based on search usage, utilizing community detection algorithms (Leiden) to synthesize shared queries into higher-level meta-clusters, reducing future search complexity.

Who it’s for

It is designed for AI agents and developers who need to perform deep, real-time search across massive local repositories and the web without the setup cost of a vector database.

Highlights

• EmbeddingDB-Free: Operates on raw files for instant search and zero information loss.
• Self-Evolving Knowledge: A living index that evolves in real-time with data changes and search patterns.
• Dual Search Modes: Offers a "DEEP" mode for rich Markdown reports and a "FAST" mode for greedy, low-latency retrieval.
• MCP Integration: Full support for the Model Context Protocol, allowing it to work with tools like Claude Desktop and Cursor IDE.
• Knowledge Compile: A beta feature for offline pre-processing to build hierarchical tree indices for boosted precision.