qmd

To search local Markdown notes and transcripts without cloud dependencies, use a local search engine that runs entirely on-device. This Skill uses a hybrid retrieval engine combining BM25, vector search, and LLM reranking to return relevant results offline.

Hybrid retrieval improves vector search results by combining BM25 keyword matching with vector embeddings and LLM-based reranking. This approach ensures both exact keyword lookups and concept-based semantic queries return highly relevant context from indexed collections.

Yes, you need Node.js version 22 or higher to run this local knowledge base search. You also need to download local models for embedding and reranking to ensure the process operates completely offline without relying on cloud services.

Yes, you can use MCP integration to search personal knowledge bases directly. MCP integration enables the Hermes Agent to access search tools natively without loading the Skill each time, streamlining collection management and context descriptions.

The best way to manage multiple collections for local search is using CLI workflows provided by the engine. You can index folders, generate embeddings, and run context descriptions across multiple project documentation and note collections efficiently.

Limitations of running LLM reranking entirely on-device include requiring sufficient local hardware resources to execute downloaded models. You must manage local installations and prerequisites manually, as the system avoids cloud dependencies for its hybrid retrieval processes.