Use case
AI Agent & Chatbot Feedback
Ask the user whether the answer helped. Capture their rating and reason with the context your team needs to investigate.
A chatbot can finish a conversation without solving the user’s problem. Collecting a rating and a reason gives your team a direct signal from the person using it.
Use Pollenate to capture the user’s own feedback on an answer or conversation, then retain the context needed to investigate it. This is distinct from using an assistant to build and manage surveys; that workflow is covered in the MCP integration guide.
Ask the user rather than guessing their satisfaction
Add thumbs up/down, a star rating, or CSAT beside the response or at the end of the task. Let the user explain what went wrong with an optional comment.
Do not generate a positive rating because the assistant believes it answered well. Record an explicit choice, or clearly distinguish any model-generated evaluation from user feedback in your own analysis.
For an existing web application, start with the React feedback component. For your own interface, call /collect when the user submits.
Keep the rating and useful context together
This request illustrates a user-provided thumbs-down on a setup answer:
curl --fail-with-body https://api.pollenate.dev/collect \
-H "Content-Type: application/json" \
-H "X-Pollenate-Key: $POLLENATE_COLLECT_KEY" \
-d '{
"inboxKey": "YOUR_INBOX_KEY",
"type": "thumbs",
"score": 0,
"comment": "The answer missed the authentication step.",
"sessionId": "example-session-id",
"context": {
"surface": "support-assistant",
"model": "YOUR_MODEL_VERSION",
"promptVersion": "setup-v3",
"intent": "api-setup"
}
}'
Use an organization key with only the collect scope, restricted to the correct brand where appropriate. /collect does not accept personal keys or OAuth tokens. The metadata field is context, not meta.
A response should return HTTP 201 before your interface shows success. Handle rejected requests and let the user retry. The developer page explains common error responses.
Store enough to investigate, not a transcript by default
A conversation or message identifier can help your application reconnect the rating to a support case. Model and prompt versions help you understand which implementation the user experienced.
Keep unnecessary personal data and credentials out of context. You do not need to send an entire conversation to record a rating. Plan how your own system will retrieve any additional context needed for investigation.
Custom metadata is retained with the event. Do not assume Pollenate automatically turns every model version or experiment into a dedicated analytics dimension.
Read the feedback before changing the prompt
Search comments for a problem such as “missing setup instructions,” then inspect the matched responses. Semantic search in Pollenate helps retrieve comments by meaning; it does not automatically cluster every failure or prove what caused it.
Use the context to reproduce the experience. Link relevant feedback to the work that will address it, and use email or webhook automation if your team needs a follow-up workflow.
After changing a prompt or model, compare similar users and tasks, and report the number of ratings alongside any percentage. A few thumbs-down responses can identify a concrete issue without being a reliable evaluation of the entire system.
Collecting feedback through MCP
The MCP server also exposes submit_feedback, subject to connection scopes and permissions. If an assistant submits on the user’s behalf, preserve what the user actually said and selected. Do not substitute inferred sentiment for a response.
For an assistant that should create, publish, and inspect your surveys, connect Pollenate through MCP. For a chatbot UI that needs a rating, create a collection key and verify the first event in your inbox.