LlamaIndex

A data framework for building LLM applications, maintained by run-llama under an MIT license. It provides data connectors, indexing and data-structuring tools, and retrieval/query interfaces, with both a high-level API (roughly five lines of code) and customisable lower-level APIs.

🔗 https://www.llamaindex.ai/ · https://github.com/run-llama/llama_index

The project has since repositioned around document agents and OCR, describing itself as “the leading document agent and OCR platform”; LlamaParse is the enterprise document parsing/extraction product, with agentic OCR across 130+ formats.

Why It Matters Here

LlamaIndex produced much of the empirical RAG-tuning literature this vault draws on — the chunk-size study, the embedding/reranker bake-off, and the long-context retrieval analysis all came out of its blog and its RetrieverEvaluator module. Its evaluation tooling is the reason those posts carry numbers rather than opinions.

Multi-Modal Abstractions

Introduced for Multimodal RAG: the OpenAIMultiModal class, a MultiModalEmbedding base class with a ClipEmbedding implementation, and MultiModalVectorIndex for indexing text and image modalities into separate vector store collections.

Related Tools

• Haystack (deepset) · LangChain — comparable orchestration frameworks
• RAGAS — evaluation
• Qdrant Vector DB · Pinecone Vector DB · Weaviate Vector DB — backing stores

Related Concepts

• RAG · Text Chunking · Multimodal RAG · Query Routing · Long-Context RAG

Articles

• Evaluating the Ideal Chunk Size for a RAG System using LlamaIndex
• Boosting RAG - Picking the Best Embedding and Reranker models
• NVIDIA Research - RAG with Long Context LLMs
• Multi-Modal RAG - Indexing And Retrieval Guide
• Hands-On RAG guide for personal data with Vespa and LLamaIndex