How GroundTruth works
Every AI-generated support answer passes through a five-stage verification pipeline before it reaches your customers. Here is exactly what happens under the hood.
Five-stage verification pipeline
GroundTruth sits between your AI agent and your customers. Every response is decomposed, verified, scored, and optionally rewritten — all in under two seconds.
Claim Extraction
Evidence Retrieval
Risk Scoring
Safe Rewrite
API Response
Claim Extraction
The AI-generated response is decomposed into atomic, independently verifiable claims. Each factual statement is isolated so it can be checked on its own against your knowledge base.
- Decomposes compound sentences into atomic claims
- Identifies factual statements vs. opinions or greetings
- Preserves context so claims remain meaningful in isolation
- Handles multi-paragraph responses with nested structures
Example
AI Draft
"You can get a full refund within 60 days of purchase, and we also offer free shipping on all orders over $50."
Claim 1
"Full refund within 60 days of purchase."
Claim 2
"Free shipping on orders over $50."
Hybrid Search
Vector Search
Semantic similarity via embeddings catches meaning even when wording differs.
Keyword Search
BM25-based retrieval catches exact terms, product names, and policy numbers.
Rank Fusion
Results from both methods are merged using Reciprocal Rank Fusion for best-of-both-worlds retrieval.
Evidence Retrieval
For each extracted claim, GroundTruth searches your knowledge base using hybrid retrieval — combining vector similarity search with keyword-based BM25 ranking — to find the most relevant evidence passages.
- Hybrid search combines semantic and keyword matching
- Reciprocal Rank Fusion merges result sets
- Top-k evidence chunks returned with relevance scores
- Works with PDFs, HTML, Markdown, and plain text sources
Risk Scoring
Each claim is classified against the retrieved evidence and assigned a verdict: Supported, Unsupported, or Needs Review. An overall risk score is computed for the entire response.
- Per-claim verdicts: Supported, Unsupported, Needs Review
- Aggregate risk score (0-100%) for the full response
- Configurable thresholds for auto-block, review, and pass
- Detailed evidence citations for every verdict
"Refund within 30 days of purchase"
"60-day refund window"
"Free shipping on orders over $50"
Before & After
Original (Risky)
"You can get a full refund within 60 days of purchase, and we also offer free shipping on all orders over $50."
Rewritten (Safe)
"You can get a full refund within 30 days of purchase. For details about shipping thresholds, please check our shipping policy page."
Safe Rewrite
When the risk score exceeds your threshold, GroundTruth automatically rewrites the response using only verified claims. Unsupported facts are corrected or replaced with soft deflections.
- Preserves the original tone and style of the response
- Replaces unsupported claims with verified facts from KB
- Uses soft deflection for claims that cannot be verified
- Configurable: auto-rewrite, flag for review, or block
API Integration
A single REST endpoint is all you need. Send the AI draft and get back a verified response with full evidence — in under two seconds.
Single endpoint
POST /api/check with the AI draft and knowledge base ID. Get back verdicts, risk score, and safe rewrite.
Sub-second latency
Optimized retrieval with FAISS indexes and parallel claim verification. Typical response in under 2 seconds.
Works with everything
Integrate with Intercom, Zendesk, LangChain, custom chatbots, and any platform that speaks REST.
Quick example
curl -X POST https://api.groundtruth.com/api/check \
-H "Authorization: Bearer hg_sk_..." \
-H "Content-Type: application/json" \
-d '{
"kb_id": "kb_abc123",
"draft": "You can return items within 60 days."
}'Ready to verify your AI answers?
Start verifying AI responses in minutes. No credit card required for your 14-day trial.