Direct API

Use LongTracer with any framework — Haystack, custom pipelines, or any code that produces strings.

Basic Usage

from longtracer import CitationVerifier

verifier = CitationVerifier()

result = verifier.verify_parallel(
    response="The Eiffel Tower is 330 meters tall and located in Berlin.",
    sources=[
        "The Eiffel Tower is a wrought-iron lattice tower in Paris, France.",
        "It stands 330 metres tall on the Champ de Mars."
    ]
)

print(result.trust_score)          # 0.5
print(result.hallucination_count)  # 1
print(result.all_supported)        # False

With Source Metadata

result = verifier.verify_parallel(
    response="...",
    sources=["chunk 1", "chunk 2"],
    source_metadata=[
        {"source": "doc.pdf", "page": 1},
        {"source": "doc.pdf", "page": 2}
    ]
)

for claim in result.claims:
    print(f"Claim: {claim['claim']}")
    print(f"  Supported: {claim['supported']}")
    print(f"  Score: {claim['score']:.3f}")
    print(f"  Source: {claim['best_source_metadata']}")
    print(f"  Hallucination: {claim['is_hallucination']}")

VerificationResult Fields

Field	Type	Description
trust_score	float	0.0–1.0, fraction of supported claims
all_supported	bool	True if all claims are supported
hallucination_count	int	Number of hallucinated claims
claims	List[Dict]	Per-claim results (see below)
flagged_claims	List[Dict]	Claims that failed verification
hallucinations	List[Dict]	Claims flagged as hallucinations
latency_stats	Dict	STS/NLI call counts and timing

Per-Claim Result Fields

Field	Type	Description
claim	str	The claim text
supported	bool	Whether the claim is supported
score	float	Final verification score
best_score	float	Best STS similarity score
entailment_score	float	NLI entailment probability
contradiction_score	float	NLI contradiction probability
nli_ran	bool	Whether NLI was executed (gated by STS)
best_source	str	Best matching source sentence
best_source_index	int	Index into sources list
best_source_metadata	Dict	Metadata of best source
is_hallucination	bool	Whether flagged as hallucination
is_meta_statement	bool	Whether it's an honest uncertainty statement

With RAG Result Dict

rag_result = {
    "answer": "...",
    "source_documents": [
        {"page_content": "...", "metadata": {"source": "doc.pdf"}}
    ]
}

result = verifier.verify_with_rag_result(rag_result)

Threshold Configuration

# Default threshold is 0.5 — adjust for stricter/looser verification
verifier = CitationVerifier(threshold=0.7)