langchain

Use create_agent with a chat model, a list of tool functions, and a system prompt. LangChain handles the ReAct reasoning loop, and you invoke the agent with a messages array to get the final response.

Load documents with a document loader, split them with RecursiveCharacterTextSplitter, embed them into a vector store like Chroma, then connect a retriever to a RetrievalQA chain. The chain retrieves relevant chunks and generates cited answers.

LangChain suits quick agents and RAG with high-level abstractions and minimal code. LangGraph fits complex stateful workflows needing cycles, fine-grained control, multi-agent systems, and human-in-the-loop patterns.

Yes, LangChain supports OpenAI, Anthropic, Google Gemini, Azure OpenAI, and local models via Ollama with a unified interface. You can swap providers by changing the model class without rewriting application logic.

Common causes include vague tool descriptions, too many tools, and missing max_iterations limits. Set max_iterations on the AgentExecutor, add error handling inside tools, and enable verbose mode or LangSmith tracing to inspect reasoning steps.

Use 512-1024 tokens for general Q&A with 10-20% overlap between chunks. Smaller chunks of 256-512 tokens suit short answers, while 1024-2048 tokens work when answers need longer context.