Local Knowledge Base Q&A System (RAG)

Objectives

• Use LangChain's PyPDFLoader or TextLoader to successfully load at least one local document and print the number of pages/chunks
• Use RecursiveCharacterTextSplitter to split the document into chunks with customizable chunk_size and chunk_overlap parameters, and verify the chunking results
• Vectorize the text chunks using an Embedding model, store them in a vector database, and implement local persistence
• Build a RetrievalQA retrieval chain that can retrieve relevant content from the knowledge base based on user questions and generate accurate answers
• Include the source documents referenced in each answer to make responses traceable back to their original sources

Hints

• Before loading a PDF, make sure to install the required dependency: `pip install pypdf`. If you encounter garbled text when loading Chinese PDFs, try using `UnstructuredPDFLoader` as an alternative
• Setting chunk_size too large will result in imprecise retrieval, while setting it too small will lose contextual meaning — it's recommended to start with 500 characters and adjust based on answer quality
• If you'd rather not use OpenAI Embeddings (which require payment), you can use `HuggingFaceEmbeddings` to load a free local model, such as `sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2`, which supports Chinese
• After setting `return_source_documents=True`, the result dictionary will contain a `source_documents` field — you can print each source's `metadata` to view page numbers and file names
• Once the vector store is built, use `vectorstore.save_local()` to persist it locally and load it next time with `FAISS.load_local()`, avoiding redundant vectorization and unnecessary API costs

Comments (0)