Llamaindex V2

LlamaIndex is the data framework for LLM applications — built around the RAG pattern (load documents → index → retrieve → generate). For agents: load 100+ data source types (PDF, CSV, database, web, Notion, Slack) via data connectors, index into any vector store (Qdrant, Weaviate, Chroma, Pinecone, etc.), query with VectorStoreIndex.as_query_engine() for semantic search + generation, and build multi-step agents with ReActAgent + tool definitions. LlamaCloud provides a hosted managed data pipeline for production-scale ingestion. 40k+ GitHub stars. Confidence is docs-derived.

Python SDK: pip install llama-index. Load documents: from llama_index.core import SimpleDirectoryReader; docs = SimpleDirectoryReader("./data").load_data(). Index: from llama_index.core import VectorStoreIndex; index = VectorStoreIndex.from_documents(docs). Query engine: query_engine = index.as_query_engine(); response = query_engine.query("What is the main theme?"); print(response.response). Streaming: query_engine = index.as_query_engine(streaming=True); streaming_response = query_engine.query("..."); streaming_response.print_response_stream(). Custom LLM: from llama_index.llms.openai import OpenAI; from llama_index.core import Settings; Settings.llm = OpenAI(model="gpt-4o"). Custom embeddings: Settings.embed_model = OpenAIEmbedding(model="text-embedding-3-small"). Vector store: from llama_index.vector_stores.qdrant import QdrantVectorStore; import qdrant_client; client = qdrant_client.QdrantClient(url="..."); vector_store = QdrantVectorStore(client=client, collection_name="docs"); storage_context = StorageContext.from_defaults(vector_store=vector_store); index = VectorStoreIndex.from_documents(docs, storage_context=storage_context). Agent: from llama_index.core.agent import ReActAgent; agent = ReActAgent.from_tools([query_engine.as_tool()], verbose=True).

No LlamaIndex API auth — pure local library. LLM provider keys (OPENAI_API_KEY, ANTHROPIC_API_KEY) set as env vars and used by LLM/embedding connectors. Vector store auth: passed to vector store client (Qdrant API key, Pinecone API key, etc.). LlamaCloud: API key for hosted data pipelines (LLAMA_CLOUD_API_KEY). No OAuth2. LlamaCloud data stays in LlamaCloud infrastructure — use self-hosted vector stores for data sovereignty.

Ingestion: synchronous for small datasets; use IngestionPipeline with workers for large-scale parallel processing. Chunking: default text splitter (512 tokens, 20% overlap) — tune for your content type. Retrieval: similarity_top_k configurable (default 2; increase for better recall at cost of context). Re-ranking: add CohereRerank or SentenceTransformerRerank post-retriever for precision improvement. LLM errors: propagated from provider (rate limits, context window exceeded). Memory: large document sets may exceed RAM — use persistent vector stores. LlamaCloud ingestion: async pipeline; monitor via LlamaCloud dashboard.

docs.llamaindex.ai covers quickstart, data connectors, vector store integrations, query engine, agents, and LlamaCloud. Getting started: pip install llama-index llama-index-llms-openai — first RAG pipeline in 5 minutes. Fully free and open-source. LlamaCloud: free tier (limited pipeline runs). GitHub: run-llama/llama_index (40k+ stars). Community via Discord (very active). Comprehensive documentation; LlamaIndex has one of the most extensive RAG framework docs available.