VER-226: [Backend] Improve Confidence Score Handling in Analysis Prompt

[nhphong]

We want the LLM to provide more reliable confidence scores in its analysis, so that the confidence scores accurately reflect the likelihood of content containing false or misleading claims.

Acceptance Criteria:

1. Revise Confidence Scoring Framework:
   • Update the prompt to include a detailed explanation of what the confidence score should represent, focusing on the degree of certainty that content contains demonstrably false or misleading claims.
   • Include a clear scoring framework with distinct categories and examples:
     • High (80-100): For content with specific false claims.
     • Medium (40-79): For misleading claims.
     • Low (1-39): For controversial opinions without false factual claims.
     • Zero (0): For content with no misleading claims.
2. Require Evidence for Non-Zero Scores:
   • Ensure the prompt requires identification and explanation of specific false claims along with credible sources for any non-zero confidence score.
3. Self-Validation Check:
   • Add a self-validation section to the prompt where the LLM must verify its scores by checking:
     • Identification of false claims.
     • Explanation of why claims are false.
     • Consistency between explanations and scores.
4. Implement Self-Review and Validation Process:
   • Introduce a structured self-review process within Stage 3, where the LLM validates confidence scores and aligns them with disinformation categories.
   • Include specific verification questions to ensure consistency and defensibility of the scores.
5. Consider Enhancements to Stage 1:
   • Explore adding self-reflection instructions to the Stage 1 prompt for better initial analysis, though confidence scoring is not part of this stage.

Tasks:

• Revise the confidence scoring section of the Stage 3 prompt with the proposed changes.
• Implement the self-validation check and self-review process in Stage 3.
• Discuss and decide on potential modifications to the Stage 1 prompt for self-reflection enhancements.
• Test the updated prompts to ensure they produce expected results and adjust as necessary.

Notes:

• This update aims to address the problem of unreliable confidence scores by providing clearer instructions and validation steps for the LLM.
• The changes are based on suggestions and discussions from the Slack conversation, focusing on improving the accuracy and reliability of confidence scores.

[linear]

VER-226 [Backend] Improve Confidence Score Handling in Analysis Prompt