build-rag-agent

# Build RAG Agent You are an expert at building AG2 RAG (Retrieval-Augmented Generation) workflows. When the user wants to build a document Q&A or knowledge-base agent: ## 1. Choose the Right Approach Ask the user: - What documents do they have? (PDFs, text files, web pages, etc.) - How large is the corpus? (Small → ChromaDB, Large → Qdrant/pgvector) - Do they need graph-based retrieval? → Graph RAG with Neo4j/FalkorDB ## 2. Basic RAG Agent (ChromaDB) ```python import os from autogen import ConversableAgent, LLMConfig from autogen.agentchat.contrib.retrieve_user_proxy_agent import RetrieveUserProxyAgent llm_config = LLMConfig( {"model": "gpt-4o-mini", "api_key": os.environ["OPENAI_API_KEY"]} ) assistant = ConversableAgent( name="assistant", system_message="You answer questions based on the provided context. If the context doesn't contain the answer, say so.", llm_config=llm_config, human_input_mode="NEVER", ) rag_proxy = RetrieveUserProxyAgent( name="rag_proxy", human_input_mode="NEVER", retrieve_config={ "task": "qa", # "qa" for Q&A, "code" for code search "docs_path": ["./docs/"], # Path to documents "collection_name": "my_collection", "chunk_token_size": 2000, "model": "gpt-4o-mini", # For embedding "get_or_create": True, # Reuse existing collection }, code_execution_config=False, ) result = await rag_proxy.a_run( assistant, message=rag_proxy.message_generator, problem="What are the main findings in the report?", ) await result.process() ``` Requires: `pip install ag2[openai,rag]` ## 3. RAG Config Options ```python retrieve_config = { # Task type "task": "qa", # "qa", "code", or "default" # Document sources "docs_path": ["./docs/", "https://example.com/page"], # Local dirs, files, or URLs # Chunking "chunk_token_size": 2000, # Tokens per chunk "chunk_mode": "multi_lines", # "multi_lines" or "one_line" "must_break_at_empty_line": True, # Break chunks at empty lines # Collection "collection_name": "my_docs", "get_or_create": True, # Reuse existing collection if available # Retrieval "n_results": 5, # Number of chunks to retrieve "distance_threshold": -1, # Max distance (-1 for no limit) # Embedding "model": "gpt-4o-mini", # Model for embeddings "embedding_model": "all-MiniLM-L6-v2", # Custom embedding model # Customization "customized_prompt": None, # Custom prompt template "customized_answer_prefix": "", # Prefix for answers "update_context": True, # Update context with new retrievals } ``` ## 4. RAG with Group Chat ```python from autogen import ConversableAgent, LLMConfig from autogen.agentchat import run_group_chat from autogen.agentchat.group.patterns import AutoPattern from autogen.agentchat.contrib.retrieve_user_proxy_agent import RetrieveUserProxyAgent llm_config = LLMConfig( {"model": "gpt-4o-mini", "api_key": os.environ["OPENAI_API_KEY"]} ) rag_proxy = RetrieveUserProxyAgent( name="rag_proxy", human_input_mode="NEVER", retrieve_config={ "task": "qa", "docs_path": ["./docs/"], "get_or_create": True, }, code_execution_config=False, description="Retrieves relevant documents. Call when facts from the knowledge base are needed.", ) analyst = ConversableAgent( name="analyst", system_message="You analyze retrieved documents and provide insights.", llm_config=llm_config, human_input_mode="NEVER", description="Analyzes documents and provides insights. Call after documents are retrieved.", ) user = ConversableAgent(name="user", llm_config=False, human_input_mode="NEVER") result = run_group_chat( pattern=AutoPattern( initial_agent=rag_proxy, agents=[rag_proxy, analyst], user_agent=user, group_manager_args={"llm_config": llm_config}, ), messages="Summarize the key risks mentioned in the documents.", max_rounds=10, ) ``` ## 5. Rules - Install RAG extras: `pip install ag2[rag]` (includes ChromaDB) - Use `get_or_create=True` to avoid re-indexing documents on every run - Set `task="code"` for code search, `"qa"` for document Q&A - `docs_path` accepts directories, file paths, and URLs - For large corpora, consider Qdrant: `pip install ag2[qdrant]` with `QdrantRetrieveUserProxyAgent` - `chunk_token_size` affects retrieval quality — smaller chunks are more precise, larger chunks have more context - Always set `code_execution_config=False` on the RAG proxy unless code execution is needed