Semantic Models

The vector model rejects a payload whose meaning is close enough to a configured set of example phrases, comparing embedding vectors instead of matching structure or exact text — useful for moderating free-text tool-call arguments or upstream responses that a schema can't constrain.

model: vector
embedding: moderator0
reject:
  - "You will never believe what happened next."
threshold: 0.94

How It Works

embedding names a configured embedding provider used to turn both the payload and each reject phrase into a vector. A payload is rejected once its cosine similarity to any reject phrase meets threshold. See Vector for the full option reference.

When to Use It

Reach for the vector model when a tool call argument or upstream response carries free text a schema can't meaningfully constrain, and the concern is content rather than structure, for example blocking a response that leaks a category of sensitive information regardless of how it's phrased. A field either validates structurally (model: json, model: avro, ...) or semantically (model: vector), not both at once — pair it with Schema Enforcement on a different field, or layer a semantic check downstream of a structural one, when a payload needs both kinds of checking.

Next Steps

• Schema Enforcement covers primitive and schema-backed models for structural validation.
• Embeddings covers the providers a vector model can reference.
• Vector covers the full reference for the model's own options.