Skip to main content

​
What counts as a hallucination?

The hallucination rate measures how often a model generates information that cannot be found in its provided inputs, such as retrieved documents, user messages, or system prompts. Quotient reports an extrinsic hallucination rate. We determine whether the model’s output is externally unsupported by the context it was given.
Extrinsic hallucinations occur when a model generates content that is not backed by any input. This is distinct from intrinsic hallucinations, where the model generates text that is self-contradictory or logically incoherent regardless of the input.We focus on extrinsic hallucination detection because this is what matters most in retrieval-augmented systems: does the model stick to the facts it was given?

Refer to How to Detect Hallucinations in Retrieval Augmented Systems: A Primer for an in-depth overview of hallucinations in augmented AI systems.

​
How Quotient detects hallucinations

  1. Segment the output into atomic claims or sentences.
  2. Cross-check every claim against all available context:
    • user_query (what the user asked)
    • documents (retrieved evidence)
    • message_history (prior conversation turns)
    • instructions (system or developer guidance)
  3. Flag unsupported claims when no context backs them up.
If a sentence cannot be traced back to any provided evidence, it is marked as a hallucination.

​
Inputs that improve detection quality

  • High-signal documents: include only the evidence actually retrieved for the answer.
  • Conversation history: pass the full multi-turn exchange so references to earlier turns can be validated.
  • Instructions: provide system prompts so the detection pass understands guardrails and policies.

​
Interpret hallucination results

  • has_hallucination: Boolean flag indicating whether we found any unsupported claims.
  • Highlighted spans: In the dashboard, statements are color-coded to show what lacked support.
  • Tag filters: Slice hallucination rate by model, feature, or customer to prioritize remediation.
Pair hallucination detection with assertions or automated tests when shipping prompt updates. A sudden spike often signals a regression in retrieval or guardrails.

​
Why monitor hallucinations?

Extrinsic hallucinations are a primary failure mode in augmented AI systems. Even when retrieval succeeds, generation can drift. Tracking this metric helps teams:
  • Catch hallucinations early in development.
  • Monitor output quality after deployment.
  • Guide prompt iteration and model fine-tuning.
Well-grounded systems typically show < 5% hallucination rate. If yours is higher, it’s often a signal that your data ingestion, retrieval pipeline, or prompting needs improvement.
Next: Document Relevance.