Skip to main content

Interface: CitationVerifierConfig

Defined in: packages/agentos/src/cognition/rag/citation/types.ts:82

Configuration for CitationVerifier.

Properties​

embedFn()​

embedFn: (texts) => Promise<number[][]>

Defined in: packages/agentos/src/cognition/rag/citation/types.ts:84

Batch embedding function: texts → embedding vectors.

Parameters​

texts​

string[]

Returns​

Promise<number[][]>


extractClaims()?​

optional extractClaims: (text) => Promise<string[]>

Defined in: packages/agentos/src/cognition/rag/citation/types.ts:95

Optional claim extractor. Falls back to sentence splitting.

Parameters​

text​

string

Returns​

Promise<string[]>


nliFn()?​

optional nliFn: (premise, hypothesis) => Promise<{ label: "neutral" | "entailment" | "contradiction"; score: number; }>

Defined in: packages/agentos/src/cognition/rag/citation/types.ts:90

Optional NLI function for contradiction detection.

Parameters​

premise​

string

hypothesis​

string

Returns​

Promise<{ label: "neutral" | "entailment" | "contradiction"; score: number; }>


pickOnly?​

optional pickOnly: boolean

Defined in: packages/agentos/src/cognition/rag/citation/types.ts:102

For a host that uses the cosine only to choose a source. With exactly one source, no embedding is made: every claim names that source with confidence: 0 and scored: false, and nliFn, when given, decides the verdict (entailment above 0.7 supported, contradiction above 0.7 contradicted, anything else unverifiable). Without nliFn the verdict is unverifiable. With more than one source the cosine ranks them as usual.


supportThreshold?​

optional supportThreshold: number

Defined in: packages/agentos/src/cognition/rag/citation/types.ts:86

Cosine similarity threshold for "supported". Default: 0.6


unverifiableThreshold?​

optional unverifiableThreshold: number

Defined in: packages/agentos/src/cognition/rag/citation/types.ts:88

Below this threshold, claim is "unverifiable". Default: 0.3