Uncertainty measures how much a model is deciding randomly between multiple ways of continuing the output, indicating the model’s confidence level in its responses.
Uncertainty is measured at both the token level and the response level:
  • Token-level uncertainty: Indicates how confident the model is about each individual token given the preceding tokens
  • Response-level uncertainty: Represents the maximum token-level uncertainty across all tokens in the model’s response
Higher uncertainty scores indicate the model is less certain about its output, which often correlates with:
  • Hallucinations
  • Made-up facts
  • Citations
  • Areas where the model is struggling with the content

​
Calculation method

Uncertainty is calculated using log probabilities from the model:
1

Token Analysis

For each token in the sequence, the model calculates its confidence in predicting that token based on all preceding tokens in the context.
2

Response Aggregation

The system identifies the highest uncertainty value across all tokens in the response to determine the overall response-level uncertainty.
3

Model Integration

The calculation leverages log probabilities from OpenAI’s Davinci models or Chat Completion models, available through both OpenAI and Azure platforms.
Uncertainty canonly be calculated with LLM integrations that provide log probabilities:

​
OpenAI

  • Any Evaluate runs created from the Galileo Playground or with pq.run(...), using the chosen model
  • Any Evaluate workflow runs using davinci-001
  • Any Observe workflows using davinci-001

​
Azure OpenAI

  • Any Evaluate runs created from the Galileo Playground or with pq.run(...), using the chosen model
  • Any Evaluate workflow runs using text-davinci-003 or text-curie-001, if available in your Azure deployment
  • Any Observe workflows using text-davinci-003 or text-curie-001, if available in your Azure deployment
To calculate the Uncertainty metric, we require having text-curie-001 or text-davinci-003 models available in your Azure environment to fetch log probabilities. For Galileo’s Guardrail metrics that rely on GPT calls (Factuality and Groundedness), we require using 0613 or above versions of gpt-3-5-turbo.

​
Optimizing your AI system

​
Addressing High Uncertainty

When responses show high uncertainty scores, your model is likely struggling with the content. To improve your system:
Identify uncertainty patterns: Analyze where in responses uncertainty spikes occur.
Enhance knowledge sources: Provide better context or retrieval results for topics with high uncertainty.
Refine prompts: Add more specific instructions or constraints for areas where the model shows uncertainty.
Consider model selection: Some models may be more confident in specific domains.

​
Best practices

Monitor Uncertainty Hotspots

Track tokens and phrases that consistently trigger high uncertainty to identify knowledge gaps.

Implement Confidence Thresholds

Set uncertainty thresholds to flag or reject responses that exceed acceptable uncertainty levels.

Compare Across Models

Evaluate how different models perform on the same inputs to identify which ones have lower uncertainty in your domain.

Combine with Factual Metrics

Use Uncertainty alongside Correctness metrics to identify correlations between model confidence and factual accuracy.
When analyzing Uncertainty, remember that some level of uncertainty is normal and even desirable in certain contexts. Very low uncertainty might indicate the model is being overly deterministic or repeating memorized patterns rather than reasoning about the content.