A local semantic search service providing hybrid retrieval and cross-encoder reranking for multi-collection knowledge bases.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install hermit-search
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install hermit-search using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is hermit-search?
hermit-search is a high-performance local search engine specifically designed to empower Openclaw Skills with semantic retrieval capabilities. It provides a privacy-first approach to knowledge management by running entirely on your local machine, utilizing CPU-optimized ONNX models for inference. The skill excels at indexing large document collections and delivering precise results through hybrid retrieval—a combination of Dense vector embeddings and Sparse BM25 indexing—followed by a sophisticated Cross-Encoder reranking step.
By integrating this skill, developers can grant their AI agents the ability to query, manage, and synchronize multiple local knowledge bases without the latency or privacy concerns associated with cloud-based search APIs. Whether you are managing code repositories, project notes, or research papers, hermit-search ensures your Openclaw Skills have the most relevant context at their fingertips.
hermit-search Use Cases
- Performing semantic search across vast local knowledge bases for precise information retrieval.
- Managing multiple distinct document collections for different projects or domain-specific research.
- Optimizing AI agent responses by providing high-quality context through hybrid Dense and Sparse retrieval.
- Automating the indexing and synchronization of local directories with advanced ignore rules for cleaner data sets.
How hermit-search Works
- The user installs the tool and downloads approximately 1GB of pre-trained embedding and reranking models.
- The hermit service is started locally, initializing an embedded Qdrant vector database and the Fastembed inference engine.
- Directories are registered as collections, where the skill recursively scans and indexes text files based on user-defined glob patterns and extension filters.
- When a search query is issued, the system performs a hybrid search using both vector embeddings and BM25 keywords.
- A Cross-Encoder reranks the top candidates to ensure the highest relevance before returning a JSON-formatted response to the calling agent.
hermit-search Setup
To get started with this skill for Openclaw Skills, follow these installation steps:
# Install hermit using uv
uv tool install git+https://github.com/xxxgqcoder/hermit.git
# Deploy the skill for agent discovery
hermit install-skills
# Download the necessary models (approx. 1GB)
hermit download
# Start the local search service
hermit start
hermit-search Data Schema & Taxonomy
The skill organizes its data and metadata within a structured local environment to support efficient Openclaw Skills operations:
| Component | Details |
|---|---|
| Data Directory | Defaults to ~/.hermit/ (configurable via HERMIT_HOME). |
| Vector Store | Uses an embedded Qdrant instance for vector and sparse data. |
| Model Stack | Jina Embeddings v2 (768-dim), Qdrant BM25, and Jina Reranker v2. |
| Collection Metadata | Tracks paths, names, glob ignore patterns, and extension exclusions. |
| Output Format | All CLI and service responses are returned as structured JSON. |
hermit-search Advanced Features
- Hybrid retrieval architecture combining jina-embeddings-v2-base-zh with BM25 for superior accuracy.
- Sophisticated reranking using jina-reranker-v2-base-multilingual to filter search results.
- Comprehensive collection management including sync status, task queues, and real-time logs.
- Advanced path filtering using glob patterns similar to .gitignore (e.g.,
**/temp/*). - Pure CPU inference via ONNX Runtime, ensuring high performance without requiring a dedicated GPU.
SKILL.md
Loading
Related Openclaw Skills
mstore McDonald's China Assistant
A comprehensive automation skill for managing McDonald's China coupons, points redemption, and delivery orders via MCP services.
Local Speech-to-Text (Whisper)
A privacy-focused, zero-cost local speech-to-text skill powered by faster-whisper for instant audio transcription.
Memory Manager
An automated system that manages the full lifecycle of an AI agent's memory through tiered storage and compression.
Skill Architect
A specialized framework for designing AI agent skills based on proven structural patterns rather than just syntax.
Bing Image Search MCP
A specialized skill enabling AI agents to programmatically search and retrieve image data from Bing via the OneKey Gateway.
Financial Analyzer
An AI-powered financial assistant for deep statement analysis, investment valuation, and automated risk assessment.
Featured*
