> ## Documentation Index
> Fetch the complete documentation index at: https://docs.wazoo.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Vector databases

> Worlds vs vector databases: similarity search against verifiable context.

## At a glance

|                | Worlds                           | Vector databases     |
| :------------- | :------------------------------- | :------------------- |
| Data model     | RDF triples                      | Embedding vectors    |
| Retrieval      | Vector + full-text + RDF filters | Nearest neighbors    |
| Relationships  | First-class graph structure      | None                 |
| Provenance     | Fact-level ledger                | None                 |
| Query language | SPARQL                           | SDK similarity calls |

## The philosophical difference

Vector databases (such as [Pinecone](https://www.pinecone.io),
[Qdrant](https://qdrant.tech), or
[pgvector](https://github.com/pgvector/pgvector)) answer "what is semantically
similar?" They index embeddings and return the nearest neighbors. That is a
useful primitive, and it is one signal among three in Worlds'
[hybrid search](/worlds/search).

Worlds stores facts as RDF triples with typed relationships. Retrieval can walk
the graph, filter by relationship, and fuse results with vector and full-text
signals. When an agent asks "Ethan's manager", a vector store returns the most
similar text, which can be another person's manager. Worlds resolves the exact
entity, then lets SPARQL traverse the actual relationship.

## When Worlds fits

* Answers depend on relationships, not just similarity: ownership, reporting
  lines, dependencies, provenance.
* You need exact answers you can verify with a deterministic query.
* Facts change and stale chunks must not contradict current state.

## When Worlds does not fit

* Your workload is pure similarity: deduplication, recommendations, anomaly
  detection over dense data.
* You have no relational structure to model.
* You are already on a vector store and similarity-only recall is sufficient.

## Coexistence

You can keep a vector index for fuzzy recall and add Worlds for the facts that
must be exact. Hybrid search inside Worlds already uses vector similarity as one
signal, so moving from a vector-only store to Worlds keeps the semantic layer
while adding graph precision.