vector model

Defines a model that rejects a message value whose meaning is semantically similar to a configured set of reject phrases, comparing embedding vectors by cosine similarity instead of matching exact text.

Feature is in Incubator

Read how to enable incubator features. Star and watch the Zilla repo for new releases!

The full message value is buffered and decoded as UTF-8 text before being embedded. The value itself is never modified — it is only accepted or rejected.

model: vector
embedding: moderator0
reject:
  - "You will never believe what happened next."
  - "I have a massive secret but I absolutely cannot tell anyone here."
threshold: 0.94
store: cache0

Configuration (* required)

model: vector

const

Specifies the model is vector.

embedding*

string

Name of a configured embedding used to turn the message value, and each configured reject phrase, into a vector. Any registered embedding provider works here — glove, openai, aws-bedrock, or ibm-watsonx-ai — not only the glove type shown in the example above.

reject*

array of string

One or more example phrases representing content to reject. Each phrase is embedded once, shared through the configured store, into a vector for comparison against incoming message values.

threshold*

number | Maximum: 1

Minimum cosine similarity, between 0 and 1, at or above which a message value is rejected.

store*

string

Name of a configured store (see memory, redis, and hazelcast) used to embed each configured reject phrase only once and share the result, rather than have every engine worker — and, with a distributed store, every replica — embed the same phrases independently.