Skip to main content

​
Overview

The Process Feedback workflow is the core automation that powers intelligent feedback triaging in GTM Feedback. It uses AI agents to automatically match incoming customer feedback to existing feature requests, creating new requests when needed, and managing the human-in-the-loop approval process.
Located at: apps/www/src/workflows/process-customer-entry/index.ts

​
Purpose

This workflow solves a critical problem in feedback management: automatically determining whether new customer feedback should be added to an existing feature request or if it represents a genuinely new request. By using semantic search and confidence scoring, it reduces manual triaging work while maintaining quality through human oversight.

​
Trigger Conditions

The workflow is triggered when:
  • A user submits feedback through the web UI
  • A Slack message is marked with a feedback reaction emoji
  • An unprocessed feedback entry exists in Redis and needs processing

​
Input Parameters

type Args = {
  customerPain: string;
  severity: "low" | "medium" | "high";
  accountId: string;
  opportunityId: string;
  links: string[];
  userId: string;
  source: "slack" | "ui";
  submitterName?: string;
  submittedBy?: string;
  feedbackId?: string; // Redis ID for cleanup
};

​
Execution Steps

The workflow implements a sophisticated three-tier confidence system:
1

Semantic Search

Uses the AI search agent to find similar feature requests based on the customer pain description.
const searchResult = await searchRequestsStep(customerPain);
const topMatch = searchResult.matches[0];
Returns matches with confidence scores (0-1) and reasoning.
2

Confidence Evaluation

Evaluates the top match against three confidence thresholds:
  • High confidence (≥0.9): Auto-add feedback to existing request
  • Medium confidence (0.8-0.9): Request human approval
  • Low confidence (<0.8): Create new feature request
These values are tuned based on semantic similarity scores from OpenAI embeddings. A confidence of 0.9+ indicates very high semantic similarity, while 0.8-0.9 represents “probably related” cases that benefit from human judgment.
3

High Confidence Path (≥0.9)

Automatically adds feedback to the matched request:
  1. Creates feedback entry with metadata (confidence, reason, match type)
  2. Auto-follows the request for the submitter
  3. Sends confirmation DM via Slack with AI-generated message
await createFeedbackForRequest({
  requestId,
  userId,
  severity,
  accountId,
  opportunityId,
  customerPain,
  links,
  metadata: {
    confidence: matchResult.confidence,
    reason: matchResult.reason,
    matchType: "existing_request",
  },
});
4

Medium Confidence Path (0.8-0.9)

Implements human-in-the-loop approval:
  1. Generates approval message using AI slack agent
  2. Creates a workflow hook for async response
  3. Sends Slack DM with “Add to Existing” and “Create New” buttons
  4. Waits for user decision
  5. Processes based on approval/decline
const hook = createHook<{
  approved: boolean;
  messageTs?: string;
  channelId?: string;
}>({ token });

await sendFeedbackMatchApprovalDm({
  token,
  slackUserId,
  message: approvalMessage,
});

const approval = await hook; // Pauses workflow until response
The workflow uses durable execution, so it can wait for user input without blocking system resources. The state is persisted and resumed when the user responds.
5

Low Confidence Path (<0.8)

Creates a new feature request:
  1. Fetches all product areas for AI assignment
  2. Uses request creation agent to generate title, description, and area assignment
  3. Creates request with auto-generated slug
  4. Stores vector embedding for future searches
  5. Creates feedback entry linked to new request
const productAreas = await getAllProductAreas();
const newRequest = await createRequestStep({
  customerPain,
  productAreas,
  metadata: {
    matchConfidence: matchResult.confidence,
    matchReason: matchResult.reason,
    matchType: "new_request",
  },
});
6

Event Tracking & Cleanup

Finalizes the workflow:
  1. Tracks analytics event based on source (UI vs Slack)
  2. Deletes unprocessed feedback from Redis cache
  3. Sends notification DM to submitter with request URL
Returns workflow result:
return {
  status: "success",
  requestId,
  matched: boolean,
  confidence: number,
};

​
Code Example

Here’s how to invoke the workflow: 
import { processCustomerFeedback } from "@/workflows/process-customer-entry";

const result = await processCustomerFeedback({
  customerPain: "Need better API rate limit visibility in dashboard",
  severity: "high",
  accountId: "acct_123",
  opportunityId: "opp_456",
  links: ["https://slack.com/archives/C123/p789"],
  userId: "user_abc",
  source: "slack",
  submitterName: "Jane Doe",
  submittedBy: "U123456", // Slack user ID
  feedbackId: "feedback_xyz", // Optional Redis key
});

console.log(result);
// {
//   status: "success",
//   requestId: "req_789",
//   matched: true,
//   confidence: 0.92
// }

​
Workflow Steps Reference

The workflow uses several reusable steps defined in /process-customer-entry/steps.ts:
Purpose: Find semantically similar feature requestsImplementation: Calls the semantic search AI agent with the customer pain descriptionReturns: Array of matches with confidence scores and reasoning
const result = await agent.search.generate({ 
  query: customerPain 
});
return { matches: result.output.matches };
Purpose: Create a new feature request using AIImplementation:
  • Uses request creation agent to generate title, description, and area assignment
  • Creates database record with auto-generated slug
  • Stores vector embedding for future semantic searches
Location: apps/www/src/workflows/process-customer-entry/steps.ts:53
Purpose: Generate contextual Slack messages using AITypes:
  • high_confidence_match: Confirmation for auto-added feedback
  • approval_request: Human-in-the-loop approval prompt
  • dm_notification: Final notification with request URL
Location: apps/www/src/workflows/process-customer-entry/steps.ts:107
Purpose: Send interactive approval request to SlackImplementation:
  • Creates Slack message with action buttons
  • Includes workflow hook token for callback
  • Returns message timestamp for later updates
Location: apps/www/src/workflows/process-customer-entry/steps.ts:131
Purpose: Clean up temporary feedback from RedisImplementation: Deletes Redis key after successful processingLocation: apps/www/src/workflows/process-customer-entry/steps.ts:10

​
Configuration

​
Environment Variables

# Required for workflow execution
SYSTEM_USER_ID=user_system      # User ID for system-created requests
SLACK_BOT_TOKEN=xoxb-...         # Slack bot token for DMs

# AI agents (configured in @feedback/ai)
OPENAI_API_KEY=sk-...           # For embeddings and agents

​
Confidence Thresholds

Adjust these constants in the workflow file: 
const AUTO_MATCH_THRESHOLD = 0.9;    // Auto-add threshold
const APPROVAL_THRESHOLD = 0.8;      // Human approval threshold
Lowering thresholds: If you decrease AUTO_MATCH_THRESHOLD, you risk more false-positive matches. If you decrease APPROVAL_THRESHOLD, you’ll create more new requests (potentially duplicates).

​
Error Handling

The workflow includes comprehensive error handling: 
try {
  // Workflow execution
} catch (error) {
  // Feedback remains in Redis for retry
  console.error("Failed to process customer feedback:", error);
  throw error; // Workflow runtime handles retry
}
Errors during optional steps (like sending DMs) are caught separately: 
try {
  await sendSlackDm({ slackUserId, message });
} catch (error) {
  console.error("Failed to send Slack DM:", error);
  // Don't fail the workflow
}

Semantic Search Agent

Learn how semantic matching works

Request Creator Agent

Understand AI-powered request generation

Workflows Concept

Deep dive into workflow architecture

Slack Integration

Configure Slack feedback collection

Build docs developers (and LLMs) love

Get started for freeTalk to us