🔍 Hallucination Auditor
Ensure model outputs are grounded in your provided data using ISR threshold validation.
⚙️ RAG-Based Architecture
1
Upload Dataset
Your knowledge base
→
2
Vectorize
ChromaDB embeddings
→
3
Query & Retrieve
Semantic search
→
4
ISR Check
Threshold validation
→
5
LLM Response
Grounded output
ISR (Information Source Retrieval) Threshold
0.75 - 1.0
Very High
Direct match - Always allow
0.55 - 0.74
High
Closely related - Allow
0.35 - 0.54
Moderate
Some relevance - Default threshold (0.40)
0.20 - 0.34
Low
Limited relevance - Block
0.0 - 0.19
Very Low
Outside scope - Block
Vector DB
ChromaDB + Cosine
Default Threshold
0.40 (Balanced)
Prevents
Hallucinations
Auditor Workbench
Threshold: 0.40
📚 Step 1: Upload Knowledge Base
File Preview