Semantic and Hybrid Row Search

Description

Semantic and Hybrid Row Search finds conceptually related rows with vector distance and can combine those results with full-text ranking through RRF.

Semantic and Hybrid Row Search is delivered through typed package contracts rather than a surface-specific shortcut. Its primary modules separate the public input or configuration shape from validation, durable state, and the operation that produces the result. That separation lets product surfaces and integrations invoke one implementation while tenant scope and existing domain rules remain in force. Errors stay bounded to the request instead of weakening the underlying data contract.

User Story

As a knowledge worker, I find prior decisions about an idea even when the original authors used different terminology. In practice, I configure or invoke the capability through a supported surface, supply only the inputs its contract accepts, and receive a result governed by the same rules as every other caller. I can inspect the resulting row, setting, event, or query state and correct my input when validation fails. I walk away with a repeatable workflow rather than a one-off client implementation.

Extension Surface

Automatable: API-level query inputs select vector, similarity, and hybrid modes with bounded result controls.

The automatable rating is evidenced by packages/dotabases/src/query/vector-search.ts, which defines the supported seam used by trusted callers or configuration. The capability can be composed wherever that contract is available, but its invariants are not replaced by tenant-authored executable code. This keeps extension honest: callers control documented inputs and policy choices while the package owns validation and safety.

Capabilities & Limits

• Capability: Executes the scoped workflow described above through the cited live implementation paths.
• Capability: Preserves typed validation and applicable tenant, schema, lifecycle, or service boundaries.
• Limit: Requires compatible stored embeddings and a query embedder. Model-provenance guards prevent silently mixing incompatible vectors. Semantic and Hybrid Row Search supports the concrete operations described here and composes them with tenant, schema, or lifecycle checks already owned by the package. It does not grant raw database access, make unsupported inputs valid, or assume that optional providers and downstream consumers exist. Callers remain responsible for permissions and meaningful configuration, and product rendering or external delivery stays outside the package unless explicitly stated.

Implementation Map

• packages/dotabases/src/query/vector-search.ts
• serializeVectorSearch
• packages/dotabases/src/query/rrf.ts
• reciprocalRankFusion
• packages/dotabases/src/services/search-by-embedding.integration.test.ts
• formatVector

Related Features

Part of the Typed Query DSL and depends on Semantic Row Embeddings.