Indexes

A VectorStoreIndex built with the default SimpleVectorStore is used in-process, then the service restarts and rebuilds the index from documents every boot, re-embedding the whole corpus.

Persist via storage_context.persist(persist_dir=...) (docstore + index_store + vector store) and reload with load_index_from_storage(StorageContext.from_defaults(persist_dir=...)). Do not re-run from_documents on every boot. For external vector stores, reconstruct the index from the existing store,…

An integrator calls VectorStoreIndex.from_documents(docs) expecting it to write to their Qdrant collection, but never passed a StorageContext, so nodes land in the in-memory SimpleVectorStore.

Pass storage_context=StorageContext.from_defaults(vector_store=QdrantVectorStore(...)) (or build via VectorStoreIndex.from_vector_store(...)). Confirm vectors actually land in Qdrant — the default in-memory store silently 'works' until restart, masking the misconfiguration.

An integrator builds a SummaryIndex over a 50k-document corpus and uses its default query engine, then is surprised by very high token cost and latency per query.

Understand the index trade-off: SummaryIndex's default mode traverses all nodes (good for small sets / full-corpus summarization), while VectorStoreIndex does top-k semantic retrieval (right for large corpora). For 50k docs use a VectorStoreIndex (or a retriever mode on the summary index), not full…