Skip to main content

Class: CitationVerifier

Defined in: packages/agentos/src/cognition/rag/citation/CitationVerifier.ts:27

Core citation verification engine.

Constructors​

Constructor​

new CitationVerifier(config): CitationVerifier

Defined in: packages/agentos/src/cognition/rag/citation/CitationVerifier.ts:34

Parameters​

config​

CitationVerifierConfig

Returns​

CitationVerifier

Methods​

extractClaims()​

extractClaims(text): Promise<string[]>

Defined in: packages/agentos/src/cognition/rag/citation/CitationVerifier.ts:142

Decompose raw text into atomic claims using the same logic verify(text, sources) uses internally.

Uses the constructor's extractClaims callback when provided, otherwise falls back to the built-in sentence splitter. Exposed publicly so callers who want to inspect or filter the claim list before verification can do so, then hand it back to verify(claims[], sources):

const claims = await verifier.extractClaims(llmText);
const filtered = claims.filter((c) => !c.startsWith('I think'));
const result = await verifier.verify(filtered, sources);

Parameters​

text​

string

Returns​

Promise<string[]>

verify()​

verify(input, sources): Promise<VerifiedResponse>

Defined in: packages/agentos/src/cognition/rag/citation/CitationVerifier.ts:66

Verify claims against provided sources.

Accepts the input in two shapes:

  • string — the raw text the verifier should decompose into atomic claims before scoring. Uses the configured extractClaims callback (e.g. an LLM-driven decomposer) or the built-in sentence splitter. Best when the input is one block of LLM-generated prose and you want the verifier to handle decomposition.

  • string[] — a list of pre-decomposed atomic claims, used as-is without any further extraction. Best when you already broke the prose into structured claims yourself, when you're verifying claims that came from a parser other than English sentence splitting, or when you want to scope verification to a specific subset of claims.

Both paths score each claim against each source via cosine similarity and (optionally) NLI contradiction detection, then return a single VerifiedResponse with per-claim verdicts.

Parameters​

input​

Either the raw text or a pre-decomposed claim list.

string | string[]

sources​

VerificationSource[]

Sources to score every claim against.

Returns​

Promise<VerifiedResponse>