LlamaIndex Index Selector

LlamaIndex (formerly GPT Index) is a framework for connecting LLMs to external data. The index type determines how your data is stored and retrieved: VectorStoreIndex uses embeddings for semantic search, SummaryIndex is optimized for summarization over all documents, KnowledgeGraphIndex extracts entity relationships, and SQLStructStoreIndex wraps relational databases for structured queries.

Use VectorStoreIndex when users ask specific questions and you want to retrieve the most relevant chunks (top-K retrieval). Use SummaryIndex when you need to answer questions that require synthesizing information across the entire document set, such as 'what are the main themes across all these reports?'. SummaryIndex reads all nodes, so it is slower and more expensive.

KnowledgeGraphIndex extracts (subject, predicate, object) triplets from your documents and stores them as a graph. It excels at multi-hop reasoning: 'What companies did Alice work for that are headquartered in Berlin?' It also makes relationships between entities explicit and queryable, which pure vector search cannot do reliably.

Yes. LlamaIndex has integrations with Pinecone, Weaviate, Qdrant, Chroma, Milvus, Elasticsearch, and many others via the VectorStore abstraction. Pass your vector store to VectorStoreIndex.from_vector_store(). This lets you persist vectors externally and scale beyond memory limits.

For frequent updates use VectorStoreIndex with an external vector store that supports upsert (Pinecone, Qdrant). You can call index.insert(document) or index.delete_ref_doc(doc_id) at runtime. SummaryIndex and KnowledgeGraphIndex are better suited for static or infrequently updated corpora because rebuilding them is expensive.