# Inbox & Training

The Inbox is your quality control center where you review, edit, and improve AI responses to continuously train and optimize your agents.

## Overview

The Inbox displays all AI interactions for review, enabling you to:

* Monitor response quality
* Correct inaccurate answers
* Identify knowledge gaps
* Train agents through feedback
* Create knowledge base articles

## Accessing the Inbox

Navigate to **Inbox** from the main menu to see recent interactions.

## Interaction List

Each interaction shows:

* **Question**: What the user asked
* **Response**: AI's answer
* **Agent**: Which agent handled it
* **Sources**: Citations used
* **Status**: Review status
* **Timestamp**: When it occurred
* **Rating**: User feedback (if provided)

## Review Statuses

| Status           | Icon | Description            |
| ---------------- | ---- | ---------------------- |
| **Not Reviewed** | ⚪    | Awaiting review        |
| **Accurate**     | ✅    | Confirmed correct      |
| **Edited**       | ✏️   | Response modified      |
| **Inaccurate**   | ❌    | Flagged as wrong       |
| **No Answer**    | ⚠️   | Agent couldn't respond |

## Reviewing Interactions

### Mark as Accurate

When response is correct:

1. Click on interaction
2. Review response and sources
3. Click **Mark as Accurate** ✅
4. Agent learns this response was good

### Edit Response

When response needs improvement:

1. Click on interaction
2. Click **Edit Response**
3. Make corrections
4. Click **Save**
5. Choose action:
   * **Update Only**: Save for records
   * **Add to Training**: Agent learns from edit
   * **Create KB Article**: Convert to knowledge base

**Example:**

```
Original Response:
"The Pro plan costs $99/month."

Edited Response:
"The Pro plan costs $99/month for up to 50 users, with
$2 per additional user. Enterprise plans with 100+ users
get custom pricing."

Action: Add to Training ✓
```

### Mark as Inaccurate

When response is wrong:

1. Click interaction
2. Click **Mark as Inaccurate** ❌
3. Provide reason:
   * Wrong information
   * Poor sources
   * Hallucination
   * Off-topic
4. Optionally provide correct answer
5. Save

This flags the interaction for investigation and improvement.

## Filtering & Search

### Filter Options

**By Status:**

* Not Reviewed
* Accurate
* Edited
* Inaccurate
* No Answer

**By Agent:**

* Select specific agent
* Compare agent performance

**By Date:**

* Today
* Last 7 days
* Last 30 days
* Custom range

**By Source:**

* Web
* Browser Extension
* Widget
* API
* Integration

**By Rating:**

* 5 stars
* 4 stars
* 3 stars or below
* No rating

### Search

Search interactions by:

* Keywords in questions
* Keywords in responses
* Topic or category
* User ID

**Example:**

```
Search: "pricing"
Results: All interactions about pricing

Search: "error" in responses
Results: All responses mentioning errors
```

## Bulk Operations

### Bulk Review

1. Select multiple interactions (checkbox)
2. Choose action:
   * Mark as Accurate
   * Mark as Inaccurate
   * Export to CSV
   * Add to eval set
3. Confirm

### Bulk Export

Export interactions for analysis:

Format options:

* CSV (spreadsheet analysis)
* JSON (programmatic processing)
* PDF (reporting)

Fields included:

* Question, Response, Agent, Sources
* Timestamp, Status, Rating
* Metadata (tokens, latency, etc.)

## Training Your Agent

### Learning from Edits

When you edit responses:

```
1. Edit saved to database
2. Original + edited pair stored
3. Agent learns from comparison
4. Future similar queries improved
```

**Example Training Cycle:**

```
Query: "How do I reset my password?"

Original Response (v1):
"You can reset your password in the settings."

Edited Response:
"To reset your password:
1. Go to Settings > Account
2. Click 'Change Password'
3. Enter current password
4. Enter new password (min 8 characters)
5. Click Save"

Next Similar Query:
Agent now provides detailed steps ✅
```

### Creating KB Articles

Convert high-quality interactions to knowledge base articles:

1. Find accurate, well-cited interaction
2. Click **Create KB Article**
3. Review and enhance:
   * Add title
   * Organize content
   * Add tags
   * Format nicely
4. Publish to KB

**When to Create KB Articles:** ✅ Frequently asked questions ✅ High-quality responses ✅ Complete, accurate information ✅ Well-cited sources ❌ One-off, specific questions ❌ Time-sensitive information

### Identifying Knowledge Gaps

Look for patterns in "No Answer" interactions:

```
Common "No Answer" Topics:
1. API v2 endpoints (15 instances)
   → Action: Add API v2 documentation

2. New feature "Project Templates" (8 instances)
   → Action: Create feature documentation

3. Pricing for non-profits (5 instances)
   → Action: Add non-profit pricing info
```

## Analytics from Inbox

### Quality Metrics

Track response quality:

* **Accuracy Rate**: % marked accurate
* **Edit Rate**: % requiring edits
* **No Answer Rate**: % with no response
* **User Satisfaction**: Average rating

### Agent Comparison

Compare agents side-by-side:

| Metric            | Agent A | Agent B |
| ----------------- | ------- | ------- |
| Accuracy          | 92%     | 87%     |
| Avg Response Time | 1.8s    | 2.3s    |
| User Rating       | 4.5     | 4.2     |
| No Answer Rate    | 5%      | 8%      |

### Topic Analysis

Identify common topics:

* Pricing (1,234 interactions)
* Technical Support (892)
* Billing (456)
* Features (332)

## Best Practices

### 1. Regular Reviews

✅ Review inbox daily or weekly ✅ Prioritize unreviewed interactions ✅ Focus on highly-rated for KB ✅ Investigate low-rated thoroughly ❌ Don't let backlog build up

### 2. Consistent Feedback

✅ Use consistent criteria for "accurate" ✅ Provide specific edit reasons ✅ Document improvement patterns ❌ Don't have multiple reviewers with different standards

### 3. Pattern Recognition

✅ Look for repeated issues ✅ Identify knowledge gaps ✅ Track trending topics ✅ Monitor seasonal patterns ❌ Don't review in isolation

### 4. Actionable Improvements

✅ Fix identified knowledge gaps ✅ Update agent instructions based on patterns ✅ Add missing data sources ✅ Create KB articles from good responses ❌ Don't just mark without improving

## Automation

### Auto-Training

Configure automatic training:

```typescript
{
  "autoTraining": {
    "enabled": true,
    "threshold": 4.5,        // Min rating for auto-training
    "requireCitation": true, // Must have sources
    "reviewRequired": false  // Skip manual review
  }
}
```

### Auto-KB Generation

Automatically create KB articles:

```typescript
{
  "autoKB": {
    "enabled": true,
    "minRating": 5,          // Only 5-star interactions
    "minOccurrence": 3,      // Asked at least 3 times
    "requireReview": true    // Admin reviews before publishing
  }
}
```

## Troubleshooting

### Inbox Not Updating

**Check:**

1. Interactions are being created
2. Filters aren't too restrictive
3. Date range includes recent activity
4. Agent filter includes active agents

### Can't Edit Responses

**Check:**

1. User has edit permissions (Manager+)
2. Interaction isn't locked
3. Agent still exists
4. Browser has connectivity

### Training Not Working

**Check:**

1. Training feature is enabled
2. Sufficient edit examples provided
3. Allow time for learning (not instant)
4. Verify similar queries improved

## Next Steps

* [Analytics Dashboard](https://github.com/thrivapp/twig-help-docs/blob/staging/monitoring/analytics-dashboard.md) - View metrics
* [Evaluation Framework](/product/monitoring/evals.md) - Automated testing
* [Performance Tuning](/product/monitoring/performance-tuning.md) - Optimize agents
* [Knowledge Base](https://github.com/thrivapp/twig-help-docs/blob/staging/knowledge-base/overview.md) - Manage KB articles


---

# Agent Instructions: Querying This Documentation

If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter:

```
GET https://help.twig.so/product/monitoring/inbox-training.md?ask=<question>
```

The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.