Skip to main content

Fact Engine

The Fact Engine verifies factual claims using deterministic methods first, without requiring LLM calls for most claims.

Features​

  • TF-IDF Semantic Similarity - No LLM needed!
  • Keyword Overlap Analysis - Fast and deterministic
  • Entity Matching - Numbers, dates, names
  • Citation Extraction - With relevance scoring
  • Negation Detection - Catch contradictions
  • LLM Fallback - Only when confidence is low

Usage​

from qwed_sdk import QWEDClient

client = QWEDClient(api_key="qwed_...")

result = client.verify_fact(
    claim="The company was founded in 2020.",
    context="Acme Corp was founded in 2020 by John Smith in San Francisco."
)

print(result.verdict)       # "SUPPORTED"
print(result.confidence)    # 0.95
print(result.citations)     # [{"sentence": "...", "relevance": 0.98}]
print(result.methods_used)  # ["semantic_similarity", "entity_matching"]

Scoring Methods​

MethodWhat It ChecksWeight
Semantic SimilarityTF-IDF cosine distance0.25
Keyword OverlapShared important words0.20
Entity MatchNumbers, dates, names0.35
Negation ConflictContradicting statements0.20

Why Deterministic?​

BEFORE (v1): 
Claim → LLM → "I think this is supported" 😕

AFTER (v2):
Claim → TF-IDF + Entity Match → SUPPORTED ✅
       (LLM only if confidence < 0.7)

No more hallucinated fact checks!