Skip to main content
Before reasoning about facts, agents must find the correct starting point. Hybrid search disambiguates natural language queries and resolves them to the correct semantic entity (IRI) before executing symbolic logic. Standard vector search retrieves semantically similar text, but agents require factual precision. If an agent searches for “Ethan’s manager”, a pure vector query might return “Gregory’s manager” because their text embeddings are nearly identical in vector space.

​
Hybrid retrieval

Worlds combines three signals for accurate retrieval. Hybrid search layers keyword matching and structural graph filters over vector embeddings, so the agent resolves the exact item it needs.

​
Reciprocal rank fusion

Results from each signal are merged using Reciprocal Rank Fusion (RRF), an algorithm that produces a single relevance ranking: score=∑d∈D160+rank(d)score = \sum_{d \in D} \frac{1}{60 + rank(d)} This lifts results ranked highly by multiple signals to the top and suppresses noise from any single signal. Use the data-plane search endpoint to perform hybrid retrieval across a world. Data-plane requests authenticate with a wzw_ world token:
Results return the matching subject and predicate IRIs plus the literal content, so an agent can bind the exact entity before running SPARQL.