rag-knowledge-assistant

What This Skill Does

The rag-knowledge-assistant is a powerful Retrieval-Augmented Generation (RAG) agent skill designed to bridge the gap between unstructured local data and intelligent AI responses. By leveraging a high-performance Chroma vector database and the BGE-M3 embedding model, this assistant indexes your local knowledge/ directory, enabling semantic search across diverse document formats including PDF, Word (.docx), Excel (.xlsx), Markdown, and TXT files. Unlike simple keyword matching, this tool understands the intent behind your query, retrieves the most contextually relevant segments from your documents, and generates human-like answers with precise citations. It includes automated intelligent chunking to ensure context integrity and supports multi-round iterative filtering to narrow down precise information from complex documentation.

Installation

To integrate this skill into your OpenClaw environment, use the command-line interface provided by the platform: clawhub install openclaw/skills/skills/aixbinge/rag-knowledge-assistant After installation, ensure your knowledge directory is properly structured. You must run the initialization scripts to generate the vector embeddings: python index_knowledge.py --knowledge-dir ./knowledge --output-dir ./vectorstore. This process populates the database so the agent can begin answering questions about your specific business documents or technical archives.

Use Cases

This skill is perfect for teams managing internal wikis, technical support engineers referencing complex API documentation, HR departments handling employee handbooks, or researchers organizing massive sets of scientific papers. It allows users to query internal policies, technical architectures, or security standards without needing to manually sift through dozens of files.

Example Prompts

1. "从知识库查：公司最新的报销政策对发票抬头有什么要求？"
2. "RAG 查询：根据产品文档，API v2.0 的鉴权接口应该如何配置 Header？"
3. "语义检索：请概括一下我们目前针对 SQL 注入攻击的防御策略。"

Tips & Limitations

To get the best results, maintain a clean and logical directory structure within your knowledge folder. If the assistant returns irrelevant results, verify the quality of your source documents; OCR-heavy PDFs may require pre-processing. Note that the assistant uses a 0.6 similarity threshold; if you feel it is missing information, you may need to adjust the similarity threshold parameter. Because this uses local vector storage, ensure you have sufficient disk space for the vectorstore directory as the index scales with your document volume.