Guide: Classification Strategies

Intentgine offers two primary ways to understand text: Resolve and Classify. Choosing the right one improves accuracy and performance.

Resolve vs. Classify

Feature	Resolve	Classify
Output	Structured Tool Call (JSON)	Label (String)
Input	Query + Tool Definitions	Text + Classes (inline or set)
Mechanism	Complex Reasoning	Categorization
Cost	1 request	1 request

Note: Both operations cost the same. Choose based on your use case, not pricing.

Classification Sets

For best performance and cost savings, use Classification Sets instead of passing classes inline.

A Classification Set is a named, stored collection of classes attached to your App — similar to how Toolsets work for Resolve. When you use a Classification Set:

Semantic caching is enabled — similar inputs return cached results.
Batch deduplication kicks in — near-identical items in a batch are grouped and only classified once.
Consistent cache keys — the content hash of the set ensures cache stability.
Extraction (optional) — automatically split compound queries into individual sub-queries and classify each one.

Creating a Classification Set

{
  "name": "Sentiment Analysis",
  "signature": "sentiment-v1",
  "classes": [
    { "label": "positive", "description": "Expresses satisfaction or happiness" },
    { "label": "negative", "description": "Expresses dissatisfaction or frustration" },
    { "label": "neutral", "description": "Neither positive nor negative" }
  ],
  "enable_extraction": false
}

Using a Classification Set

Reference it by signature in your API calls:

{
  "data": "The food was okay but the service was slow.",
  "classification_set": "sentiment-v1"
}

Context

You can provide optional context to help the LLM make better classification decisions. This is useful when the same text could be classified differently depending on the situation.

{
  "data": "It's running hot",
  "classification_set": "support-v1",
  "context": "Customer is reporting an issue with their laptop"
}

The context field is limited to 1,000 characters.

You can still pass inline classes for ad-hoc classification — this uses exact-match caching only, as before.

Extraction

Classification sets can optionally enable automatic extraction. When enabled, the LLM can detect compound queries and extract them into individual sub-queries, then classify each one. See the Classification Extraction Guide for details.

Use Cases

When to use Resolve

You need to take an action based on input.
You need to extract parameters (dates, names, numbers).
The input is an instruction or command.
Example: “Book a flight to Paris on Friday” -> book_flight(dest="Paris", date="Friday")

When to use Classify

You need to route a message.
You want to know the sentiment.
You need to tag data for analytics.
Example: “This flight was terrible” -> sentiment: negative
Example: “I need a refund” -> department: billing

Pipeline Pattern

A common pattern is to use Classify to route a request to the correct specific agent, then use Resolve within that agent context.

Router: Classify “I need a refund” -> Label: billing.
Handler: Load billing tools.
Resolver: Resolve “I need a refund” using billing tools -> create_refund_request().

This saves tokens because the Resolver only sees relevant billing tools, not every tool in your system.