Create Knowledge Base

Kafka source configuration for the knowledge base pipeline.

Exactly one of topic or topicPattern must be set:

• topic — exact Kafka topic name (the FE wizard's single-topic picker emits this). The service layer composes ^re.escape(topic)$ for the Java --topicMatcherRegex CLI arg, so the Java side keeps the regex-only contract.
• topicPattern — escape hatch for tenants who genuinely need a multi-topic regex (e.g. fan-in of a sharded topic family). Not exposed by the FE wizard today — programmatic API callers only.

The XOR rule is enforced by _topic_xor below: empty payload (neither set) and over-specified payload (both set) both 422.

Text template configuration for composing the embedding input from record fields.

Embedding-side configuration for a knowledge base.

Carries a reference to a saved AgentLlmConnection plus optional per-KB overrides of the connection's defaults. Credentials, provider, and the default model live on the connection — there is no inline apiKey / baseUrl path on a KB. The deploy-time resolver (resolve_kb_connections) reads the referenced connection's embedding defaults block, applies any per-KB overrides, and stamps the resulting bundle into the Flink CLI JSON.

The referenced connection must carry 'embedding' in its capabilities list and have a non-empty embedding.model — the validator inside resolve_saved_embedding_credentials raises 422 at deploy time otherwise so the FE error surfaces the cause without waiting for a Java runtime 4xx.

Vector store configuration for the knowledge base.

References a lightweight AgentVectorStoreConnection by id. The BE resolves the connection's credentials (apiKey + endpoint) at deploy time. There is no inline-credentials path — KBs must reference a saved connection (the FE picker is the only authoring surface).

Processing configuration for the knowledge base pipeline.