Improve CSAT evaluator prompt to better detect explicit user dissatisfaction#4998
Open
imatiach-msft wants to merge 1 commit intomainfrom
Open
Improve CSAT evaluator prompt to better detect explicit user dissatisfaction#4998imatiach-msft wants to merge 1 commit intomainfrom
imatiach-msft wants to merge 1 commit intomainfrom
Conversation
Test Results for assets-test68 tests 68 ✅ 2s ⏱️ Results for commit 36b60ea. ♻️ This comment has been updated with latest results. |
…faction Add explicit rules to both single-turn and multi-turn CSAT evaluator prompts for handling user dissatisfaction signals: - Explicit DSAT signals (e.g. 'that didn't help!') MUST score 1 or 2 - Unresolved core requests cap at score 3 even with good tone - Multi-turn: trailing user messages without agent response scored accordingly - Professional tone does not compensate for confirmed user dissatisfaction Tested with gpt-5.2 judge model: - Multi-turn WITH DSAT: Score 2 (correctly detects dissatisfaction) - Multi-turn WITHOUT DSAT: Score 3 (neutral, unresolved) - Single-turn baseline (no DSAT): Score 4 (unchanged for happy path) Bug #5243079 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
imatiach-msft
force-pushed
the
ilmat/csat-dsat-detection
branch
from
1fc854f to
36b60ea
Compare
vebudumu
approved these changes
May 5, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Fixes Bug #5243079 - CSAT evaluator scores 3 (Neutral) when user explicitly expresses dissatisfaction, instead of scoring 2 (Dissatisfied).
Problem
When the CSAT evaluator receives a conversation where the user explicitly states the agent response was unhelpful, the model gives too much credit for the agent polite tone and alternative suggestions, inflating the score to 3 instead of reflecting the user actual dissatisfaction.
This happens specifically when the conversation is passed as flattened text in the query field (which is how the production pipeline calls the single-turn evaluator).
Changes
Added two rules to the IMPORTANT CONSIDERATIONS section of both customer_satisfaction.prompty and customer_satisfaction_multi_turn.prompty:
Test Results (gpt-5.2, temperature=0, 5 runs each)
With DSAT present in query (text format) - the bug scenario
Without DSAT (query = just user question, no follow-up)
The unresolved core requests rule prevents inflated scores (4) when the agent could not fulfill the request.
With full JSON conversation (structured format)
When conversation is passed as structured JSON, the model already handles DSAT correctly - no regression.
Test Conversation (from production trace)
User: hi
Assistant: Hi - what can I help you with?
User: whats the weather in New York?
Assistant: I cant see live weather data from here... [offers alternatives]
User: that didnt help!
Production scored this as 3 (Neutral). With our change, it correctly scores 2 (Dissatisfied).