Vectorless Database

Vectorless database is an informal label, not a single database engine or settled standard. It describes retrieval architectures that answer queries without a vector index. They may use hierarchical trees, lexical indexes, symbolic relationships, or an LLM that chooses which sections to inspect.

The design starts from a different premise than vector search. Instead of compressing a document into embeddings and ranking chunks by geometric similarity, it preserves explicit structure such as headings, page order, tables, references, and parent-child relationships. A reasoning model can then select a branch, read the relevant section, follow a cross-reference, and ask for more context when the first passage is insufficient.

PageIndex is a prominent example of this pattern. It converts a document into a hierarchical JSON tree with titles, summaries, metadata, and child nodes, then uses model-guided navigation at query time. This can work well for financial reports, legal filings, and manuals where hierarchy matters. Vector search remains useful for large, unstructured collections, and hybrid systems can use both approaches.