LlamaIndex
Définition
LlamaIndex focuses on connecting LLMs to your data: ingestion, indexing, and querying. It provides flexible RAG pipelines, multiple index types, and evaluation tools.
Il complète LangChain: LlamaIndex emphasizes the data layer (documents, embeddings, vector stores, indexing strategies). Utilisez-le quand your priority is robust RAG over your own docs, APIs, or databases, with control over chunking, récupération, and synthesis. Also supports agents and query engines.
Comment ça fonctionne
Charger des données depuis des documents, APIs ou bases de données dans un format de document unifié. Construire des index : index vectoriel (embeddings + vector store), keyword index, or hybrid; you choose node parsers (chunking), embedding model, and index type. Query engines run récupération (optionally with reranking) and then synthesis (the LLM answers from retrieved nodes). You can customize retrievers, node parsers, and response synthesis (par ex. tree summarization, simple concatenation). Evaluation tools (par ex. faithfulness, relevance) help tune chunking and récupération for production RAG. Agents can use LlamaIndex query engines as tools inside LangChain or native agent loops.
Cas d'utilisation
LlamaIndex fits when you need flexible RAG indexing, query engines, and evaluation over your own data and APIs.
- RAG and document Q&A with flexible indexing and query engines
- Connecting LLMs to internal data (docs, APIs, databases)
- Evaluating and tuning récupération and synthesis for production RAG