Skip to main content
You are here: AI Lab

Fine tuning

How to Fine-Tune Your Engine

Introduction: What is Fine-Tuning?

Fine-tuning is the process of improving your AI Engine after it has been launched by using real user feedback. It's the "continuous improvement" loop of your AI strategy.

When your agent is live, users can provide feedback on responses—thumbs up/down, star ratings, or text comments. Fine-tuning allows you to collect this feedback, review it, and use it to create improved versions of your Engine.

Think of it as:

  • Your Engine = The first draft of your AI's brain

  • Fine-tuning = Learning from mistakes and successes to create a smarter, better version

The Fine-Tuning Workflow

text

Launch Engine → Collect Feedback → Review Feedback → Create Fine-Tuning Job → New Engine Version Created

Prerequisites

Before you can fine-tune an engine, you need:

  • A published Engine that has been live with agents collecting feedback (Learn how to create AI engines)

  • User feedback collected on that engine's responses (star ratings, text comments, etc.)

  • Access to AI Labs → Fine Tuning section

Step-by-Step: Fine-Tuning Your Engine

Step 1: Access the Fine Tuning Section

  1. From the main navigation sidebar, click AI Labs .

  2. Click the sub-menu item Fine Tuning .

  3. You will see the Fine Tuning dashboard with a list of existing jobs and a + Start Tuning button at the top left.

Step 2: Start a New Fine-Tuning Job

  1. Click + Start Tuning .

  2. The Create Fine Tuning Job panel will open.

Step 3: Configure Job Basics

3.1. Title

  • Enter a descriptive name for this tuning job (e.g., "Q2 Poll Engine Improvements" or "Formal Tone Fixes").

3.2. Select Engine

  • Click the dropdown and choose the Engine you want to improve. This will be the engine that has been collecting feedback.

3.3. Select Engine Version

  • Choose which version of the engine you want to base this tuning on.

Important Note on Versions

Feedback is collected at the engine version level, not the engine level. This means:

  • If you have "Version 1" and "Version 2" of an engine, each has its own feedback history.

  • When selecting a version here, you are only tuning based on feedback received for that specific version.

  • This allows you to compare performance across versions and improve incrementally.

3.4. Date Filter

  • Select the time period for feedback you want to review:

    • Past Week

    • Past Month (default)

    • Past 3 Months

    • Custom Range

Step 4: Review and Select Feedback

This is the most critical step. You will see a table of Feedback Selection showing all feedback collected during the selected period.

Feedback Table Columns:

Column

Description

Engine Title

Which engine generated the response

Engine Version

Which version was used

Feedback Type

How the user responded (Stars, Text, Manual Edits, Regenerate)

Feedback Text

Any comments the user left

Understanding Feedback Types:

Type

Description

Stars

User rating (1-5 stars). Higher is better.

Text

User wrote a comment (e.g., "Too informal, should be more professional").

Manual Edits

A human editor manually corrected the AI's response. This is high-value feedback. This is normally used in Polls and agents which can have human in the loop

Regenerate

User clicked "regenerate" to get a different response (implies dissatisfaction).

Selecting Feedback for Tuning:

  • Browse through the feedback items.

  • Check the checkbox next to each piece of feedback you want to include in this tuning job.

  • Tip: Focus on:

    • Low-star ratings (1-2 stars)

    • Text comments with specific guidance

    • Manual edits showing correct responses

    • Patterns, multiple users complaining about the same issue

Step 5: Create the Job

  1. Once you have selected all relevant feedback, click Create Job at the bottom.

  2. You will be returned to the Fine Tuning dashboard, where your new job will appear in the list with a Processing status.

Step 6: Job Completion and New Version Creation

Once a job status shows as Completed or Published, the system has:

  1. Analyzed all the feedback you selected

  2. Updated the engine's guidelines based on that feedback

  3. Automatically created a new version of your engine

What happens next:

  • The new version appears in the engine selection dropdowns with the next version number (e.g., "Version 2").

  • The original version remains available and unchanged.

  • You can now use this new version in any of your agents.

Note: There is no separate dashboard to review the tuned job details. The completion of the job is indicated by the status change, and the result is a new, improved engine version ready for use.

Viewing Feedback Logs

At any time, you can view raw feedback for any engine:

  1. Go to Fine Tuning dashboard.

  2. Click on an engine or job to see its Feedback Logs.

  3. Use Filters to narrow by:

    • Feedback Type (Stars, Text, Manual Edits, Regenerate)

    • Date range

    • Rating value

The logs show:

  • Time of feedback

  • Engine + Version used

  • Feedback Type

  • Reason / Comment (if any)

Best Practices for Fine-Tuning

1. Focus on Quality Over Quantity

  • 10 detailed text comments are more valuable than 100 star ratings.

  • Prioritize feedback that includes specific guidance ("too formal," "add bullet points," "wrong tone").

2. Look for Patterns

  • If multiple users rate responses poorly on formality, focus your tuning on tone guidelines.

  • If users keep asking for more concise answers, adjust your Engine's structure guidelines.

3. Include Manual Edits

  • When editors manually correct AI responses, this is gold. It shows exactly what the right answer should be.

4. Tune Regularly

  • Schedule monthly or quarterly tuning sessions to continuously improve.

  • Each tuned version should be smarter than the last.

5. Version Control

  • Each tuning job creates a new version, allowing you to roll back if needed and compare performance across versions.

Troubleshooting

Issue

Likely Cause

Solution

No feedback appears

Engine hasn't collected any feedback yet

Ensure feedback collection is enabled on your agents. Wait for user interactions.

Only stars, no text

Users are rating but not commenting

Consider prompting for comments after low ratings.

Job status stuck on "Processing"

Large volume of feedback

Wait a few minutes. If still stuck, contact support.

Can't select an engine

Engine may not have any feedback yet

Check date filter. Expand to longer time period.

New version seems worse

Poor feedback selection

Review which feedback you included and try again with more focused selections.

Fine-tuning transforms your AI from good to great. By systematically learning from real user interactions, you create Engines that truly understand your audience and deliver exceptional experiences.