modelscope/sirchmunk

What it solves

Traditional retrieval systems rely on static vector embeddings, which are expensive to build, become stale quickly, and lose context from the original data. Sirchmunk removes the need for vector databases and complex preprocessing pipelines, letting you search and chat directly over raw files with zero information loss. It also keeps the index alive: as your data changes, the system evolves with it, so insights stay current without batch re-indexing.

How it works

Sirchmunk works straight from raw data using a framework called LENS (Latent Evidence Navigation and Search). Instead of converting everything into vectors, it performs agentic search with multiple parallel routes—lexical, entity, directory, structure, and cross-document topic maps—then fuses results using confidence-weighted scoring. It uses large language models (LLMs) only when needed to keep costs low, and a tight token budget keeps each query efficient. The system also builds knowledge clusters that self-evolve over time through a four-phase cycle: connect & merge, edge refresh, meta-cluster detection, and global update. This makes search both deterministic and context-aware, with answers linked directly to source evidence.

Who it’s for

Sirchmunk is built for developers and AI agent builders who need fast, real-time retrieval from large local repositories or web-scale data without the overhead of vector infrastructure. It’s especially useful for anyone integrating search into LLM-powered applications—via Python SDK, CLI, HTTP API, or the Model Context Protocol (MCP)—and for researchers exploring retrieval evaluation. It also offers a web UI for interactive exploration and chat.

Highlights

• EmbeddingDB-free: No vector embeddings, no vector database; search directly over raw files without preprocessing.
• Self-evolving knowledge base: The index and knowledge clusters update automatically as data changes, staying fresh without manual re-indexing.
• Budgeted evidence exploration: LENS balances thoroughness and cost by spending tokens only where needed, with configurable DEEP and FAST search modes.
• Real-time and scalable: Handles massive datasets with extremely low RAM/CPU consumption and native elastic support.
• Agent-ready: Full MCP server support, OpenAI-compatible LLM integration, streaming APIs, and compatibility with tools like OpenClaw.
• Developer-friendly: Python SDK, CLI commands (init, search, serve, web, mcp), Web UI, Docker images, and a research benchmark framework.