Skip to main content

How to Optimize Tagging Agents

Tagging Agents automatically classify memories using AI. Well-designed tagging agents can help with more accurate tagging and meaningful analysis.

This guide explains how to structure Tagging Agents, write effective tag descriptions, and avoid common mistakes.

Key Principles

1. One clear purpose per tagging agent​

Each Tagging Agent should answer one specific question.

Good examples:

  • What is this feedback about? → Experience Topics
  • What went wrong? → Pain Points
  • How does the customer feel? → Sentiment

Avoid mixing multiple goals into one tagging agent. This reduces accuracy and makes results harder to interpret.

2. Keep tagging agents focused and small​

A tagging agent should typically contain 5–10 tags.

Too many tags:

  • reduce classification accuracy
  • create overlap between tags
  • make results harder to analyze

3. Separate tagging concerns​

Use multiple tagging agents instead of one complex tagging agent.

Recommended structure:

  • Tagging Agent 1: Experience Topics
  • Tagging Agent 2: Pain Points
  • Tagging Agent 3: Sentiment
  • Tagging Agent 4: Persona

This separation consistently leads to better results.

4. Scope tagging agents to relevant data​

Tagging Agents should run only on relevant memories. If a tagging agent runs on everything, it will be less accurate.

Use search filters (e.g. source, type, keywords) to:

  • reduce the number of memories processed
  • improve classification precision
  • reduce cost and processing time

Example:

Instead of running a tagging agent on all data, scope it to:

  • only NPS responses
  • only support conversations
  • only product reviews

5. Start simple and iterate​

Do not try to design a perfect system upfront.

Start with a small number of tagging agents and tags, then improve based on:

  • how tags are used in analysis
  • which questions users ask in chat
  • which tags are rarely used

Writing Effective Tag Descriptions

Tag descriptions are critical. Most tagging accuracy issues come from poor tag descriptions - not from the model.

Good descriptions include:​

1. Clear definition​

Explain what the tag represents.

Example

Delivery delay
Feedback about late deliveries, missed delivery windows, or orders arriving later than expected.

2. Trigger phrases​

Provide examples the AI can match.

Trigger phrases: late delivery, arrived late, delayed shipment, delivery took too long

3. Exclusions​

Clarify what should not be tagged.

Not for: damaged items, wrong items, or missing products

Complete example​

Payment failure

Feedback about issues completing a payment during checkout.

Trigger phrases: payment failed, card declined, transaction error, could not pay

Not for: pricing complaints or refund requests

Common mistakes​

Too vague​

❌ "Issues with delivery" → unclear and overlaps with many cases

No trigger phrases​

❌ No examples provided → reduces classification accuracy

No exclusions​

❌ No boundaries defined → leads to overlapping tags

Structuring Tagging Agents

Example setup​

Tagging Agent 1: Experience Topics​

Purpose: What part of the journey is discussed?

Tags:

  • Checkout
  • Delivery
  • Returns
  • Customer Support
  • Product Quality

Tagging Agent 2: Pain Points​

Purpose: What went wrong?

Tags:

  • Delivery delay
  • Payment failure
  • Missing information
  • Technical issue

Tagging Agent 3: Sentiment​

Purpose: How does the customer feel?

Tags:

  • Positive
  • Neutral
  • Negative

Tagging Agent 4: Persona​

Purpose: Who is the customer?

Tags:

  • New Customer
  • Returning Customer
  • Loyalty Member

Testing Tagging Agents

Before activating a tagging agent, test it using the Playground:

  • validate that tags are applied correctly
  • identify ambiguous tags
  • refine descriptions and trigger phrases

A simple workflow:

  1. Add example memories
  2. Run the Tagging Agent
  3. Check assigned tags
  4. Adjust descriptions if needed

Common Mistakes

1. Too many tags​

Some teamspaces define hundreds of tags, but only a small subset is actually used in analysis.

For example, we encountered once a teamspace with 170 tags but only 15 had ever been used in chats.

This indicates unnecessary complexity.

2. Mixing segmentation and insights​

Combining tags like:

  • Country
  • Persona
  • Delivery issue

in one tagging agent leads to:

  • worse accuracy
  • harder analysis

Always separate these concerns.

3. Overly granular tags​

Examples:

  • specific product SKUs
  • detailed technical attributes
  • long demographic breakdowns

These rarely improve insights and often reduce tagging quality.

4. Missing or weak descriptions​

Tags without:

  • clear definitions
  • trigger phrases
  • exclusions

lead to inconsistent results.

5. Too many tagging agents​

Some teamspaces define many tagging agents (e.g. 8–9), but only a few are actually used in chats.

Focus on a small number of high-quality tagging agents instead.

Optimization Checklist

Use this checklist to improve your setup:

  • Each tagging agent has a single clear purpose
  • Tagging Agents contain 5–10 tags
  • Segmentation and insights are separated
  • Tagging Agents are scoped using search filters
  • Each tag has:
    • clear definition
    • trigger phrases
    • exclusions
  • Tagging Agents are tested in the Playground
  • Unused tags are removed regularly

Summary

Effective Tagging Agents are:

  • focused (one purpose per tagging agent)
  • simple (limited number of tags)
  • well-defined (clear descriptions)
  • scoped (only relevant data)
  • iterative (improved over time)

The goal is not to classify everything.

The goal is to enable accurate, actionable insights from your data.