用 Dify 搭建 RAG 文档问答系统

上传文档，构建向量知识库，用户提问时检索精确答案并附引用来源。

Legal teams reviewing contracts, HR departments with policy documents, technical teams with API documentation.

Dify account, documents to upload, OpenAI API key for embeddings

工作流图

Create Knowledge Base

Configure Retrieval

Create Q&A App

Add Citation Display

Deploy and Monitor

1
Create Knowledge Base

In Dify, create a new Knowledge Base. Upload your documents (PDF, DOCX, TXT, MD). Set chunking to 300 tokens with 50 token overlap. Choose the embedding model (OpenAI text-embedding-3-small is cost-effective).
2
Configure Retrieval

Set retrieval mode to 'semantic search' with top-K=5. Enable reranking if available (Cohere reranker improves accuracy significantly). Set a similarity threshold of 0.7 to filter irrelevant chunks.
3
Create Q&A App

Create a new Completion App. Connect your knowledge base. Set the prompt: 'Answer the user's question using ONLY the provided context. If the answer is not in the context, say I cannot find this information. Always cite the source document and page number.'
4
Add Citation Display

Enable Dify's citation feature. When the model generates an answer, it will show which chunks were used. Configure the UI to display source filenames and relevant passages.
5
Deploy and Monitor

Deploy as a web app or API. Set up logging to track: query volume, answer accuracy (user feedback), and unanswered questions (to identify knowledge gaps). Add new documents based on gaps found.