Custom Dimensions

Overview

Dimension Labs provides two types of enrichment:

• Standard Dimensions — A comprehensive set of analytical tasks available out of the box in every partner organization. These dimensions (such as Reason, Category, Predicted Rating, and Summary) are designed to work across most textual data sources with no configuration required.

  • Referred to in the platform as "System Prompts"

• Custom Dimensions — Enrichment tasks that you define to align with your specific business needs, terminology, and research objectives. Custom prompts are how you create custom dimensions.

  • Referred to in the platform as "Custom Prompts" or "Prompts"

Standard dimensions give you a strong analytical foundation from day one. Custom prompts extend that foundation by letting you extract the specific signals that matter to your organization — whether that's detecting product friction in support conversations, identifying churn intent in survey responses, or categorizing feedback by your internal taxonomy.

💡

Custom prompts are the most significant aspect of customization before you begin engaging with the AI Agent to explore your data, answer business questions, and produce knowledge from your underlying data.

---

What Is a Custom Prompt?

A custom prompt is an analytical task you define that tells the AI exactly what to look for in your data. Each custom prompt runs against every transcript in an integration and outputs a labeled value — creating a new column (dimension) in your enriched dataset.

Think of it this way: if a standard dimension like Reason asks "What is this conversation about?" — a custom prompt lets you ask anything specific to your business, such as:

• "Did the customer mention a specific product feature?"
• "Is there a billing dispute in this conversation?"
• "What stage of the customer journey does this interaction represent?"
• "How would you rate the quality of the agent's response?"

Each of these questions becomes a dimension in your data — a new column you can filter, cross-tabulate, and visualize alongside every other dimension in the platform.

---

How Custom Prompts Relate to Enrichment

Enrichment is the process of extracting meaning from raw text and applying it as structured labels. Each enrichment task uses a prompt to analyze individual transcripts with a large language model and output a labeled value.

Custom prompts follow the same enrichment process as standard dimensions. The difference is that you control the question, the output format, and the classification logic. A custom prompt can output:

• A dynamic label generated from the text (e.g., a reason or summary in the customer's own language)
• A predefined category selected from a list you define (e.g., Billing, Technical Issue, Product Inquiry)
• A boolean value indicating whether a condition is present (e.g., True if friction is detected)
• A score or rating on a defined scale (e.g., a 1–5 CSAT prediction)
• A verbatim quote extracted from the customer's messages

These output types can be combined within a single custom prompt to create a rich set of related fields that work together.

---

Prompt Functions

Custom prompts are built using prompt functions — structured definitions that contain one or more fields, each representing a single analytical task. A prompt function groups related fields together so they are processed as a unit.

A prompt function includes:

• Context — A brief description of the data being analyzed, including who the incoming and outgoing users are and what the organization does. This gives the AI the background it needs to interpret the data accurately.
• Fields (Parameters) — The individual analytical tasks within the prompt function. Each field has a name, a description of what to extract, and a defined output type.

Example: A Simple Prompt Function

Imagine you run an e-commerce company and want to understand why customers contact support. You could define a prompt function with the following fields:

When this prompt function runs, every transcript in your integration is analyzed and each field produces a labeled output — giving you six new dimensions of structured data to work with.

---

Output Types

Custom prompt fields support several output types, each suited to different kinds of analysis.

Dynamic Labels

The AI generates a label based on what it finds in the text. No predefined options are needed — the output reflects the actual content of the conversation.

🎯

Example: A contact_reason field might output labels like order tracking inquiry, refund request, product availability question, or shipping delay complaint — all generated organically from the data.

Dynamic labels are especially powerful for discovery. Because the AI isn't constrained to a fixed list, it can surface themes and patterns you didn't anticipate. These organic labels can then be consolidated into broader themes using data mapping.

Predefined Categories

The AI selects from a list of categories you define. Use this when you have an established taxonomy or need consistent groupings for reporting.

🎯

Example: An engagement_category field with predefined options like Billing & Payments, Account Access, Product Setup, Returns & Refunds, and Unspecified.

Every category field should include a catch-all option (such as Unspecified or Other) for cases where no category is a strong match.

Boolean Fields

A simple True/False output indicating whether a specific condition exists in the transcript.

🎯

Example: An issue_detected field that outputs True if the customer reports a problem or expresses frustration, and False otherwise.

Boolean fields are often used as gates for other fields — for instance, issue_reason and issue_verbatim only need to produce output when issue_detected is True.

Scores and Ratings

A numerical value on a defined scale, such as a predicted CSAT score (1–5) or a satisfaction rating (1–10).

🎯

Example: A csat field on a 1–5 scale where 1 = Very Dissatisfied and 5 = Very Satisfied.

When there is not enough signal in the transcript to assign a score with confidence, the field outputs a null value rather than guessing — keeping your data clean for aggregation.

Verbatim Quotes

The AI extracts a direct quote from the customer's messages that is most relevant to a given topic.

🎯

Example: An issue_verbatim field that quotes what the customer said about their problem in fewer than 30 words, pulling only from incoming (customer) messages.

Verbatim fields preserve the customer's original language, making them valuable for qualitative review and stakeholder reporting.

---

Field Blocks: A Repeatable Pattern

Custom prompts are most effective when related fields are grouped into consistent field blocks. A standard block follows this pattern:

This pattern repeats for each type of signal you want to track — issues, friction points, product mentions, feature requests, and so on. Using consistent blocks makes your dashboards predictable and your data easy to work with.

---

Conditional Logic and Null Values

Fields within a block are often conditional — they depend on a parent boolean field. For example, issue_reason only needs to produce output when issue_detected is True. When the condition is not met, the field outputs a null value.

This is an important design principle: null values (not placeholder strings like "N/A" or "none") keep your data clean for filtering, aggregation, and visualization. When you filter your dashboard to records where issue_detected is True, every record in view has a meaningful issue_reason and issue_category.

---

Custom Prompts vs. Standard Dimensions

Standard and custom dimensions work together. Standard dimensions provide a universal baseline — Reason, Category, Predicted Rating, Summary, and more — while custom prompts layer on the specific signals your teams need to make decisions.

⚡

Best Practice: Start with standard dimensions to establish a baseline understanding of your data, then build custom prompts to go deeper into the areas that matter most to your business.

---

Common Use Cases

Custom prompts are flexible enough to support a wide range of analytical objectives. Here are some common patterns:

Product and Feature Mentions

Detect when customers mention specific products, features, or services. Each product or feature is defined as its own boolean field so that co-occurring mentions are captured independently.

Issue and Friction Detection

Identify when customers experience problems, what those problems are, and how they should be categorized for resolution. Combine boolean detection with reason labels, verbatim quotes, and root-cause categories.

Customer Journey Classification

Determine what stage of the customer lifecycle an interaction represents — onboarding, active use, renewal consideration, or churn risk — and track how conversation topics shift across stages.

Satisfaction and Quality Scoring

Predict customer satisfaction (CSAT), agent helpfulness, or support quality on defined scales. These scores can be compared across segments, time periods, and categories to surface trends.

Domain-Specific Signals

Detect signals unique to your industry or operation — compliance mentions in financial services, safety concerns in healthcare, revenue opportunities in sales conversations, or competitive mentions in product feedback.

---

Getting Started

Custom prompts are specific to an integration within your partner organization. To create custom prompts:

1. Identify your objectives — What business questions do you need to answer? What decisions will the enriched data support?
2. Define your fields — Decide what information to extract: topics, issues, products, scores, verbatims, or any combination.
3. Align to your taxonomy — If your organization has existing categories, product lines, or classification schemes, incorporate them into your predefined category fields.
4. Test and refine — Sample the output on real data and adjust field descriptions, category lists, and conditional logic as needed.

Your Dimension Labs implementation team will work with you to design prompt functions that align with your data sources and analytical goals. Custom prompts can also be created and edited directly through the platform's no-code prompt interface.

Pro Tip: When designing custom prompts, think about how fields will be used together. Cross-tabulating an engagement_category (what the customer is talking about) against an issue_category (what needs to change to resolve it) can reveal patterns that neither field surfaces on its own.