Dimensions

What a Dimension is

Dimensions are the typed fields Dimension Labs extracts from every conversation. What they are, the types, and how to read them.

A Dimension is a structured field extracted from a conversation. Sentiment is one question you can ask of text. Dimensions let you ask many: why did this customer get in touch, did they mention a competitor, how much effort did it take, what would they rate it, what exactly did they say about the delivery.

enrichment · one record

Raw text

We're running a 60-day parallel evaluation of another vendor before we sign the renewal. Honestly, we were promised SSO in Q2 and it still isn't here, and this ticket has been open for three weeks.

Support ticket · enterprise account · ARR joined

Dimensions written to the record

Competitor Mentionedyes · evaluation underway
Churn & Cancellation Riskhigh
Lifecycle Stagerenewal
Urgency & Deadline60 days
Expectation Gaproadmap promise missed
Feature RequestSSO
Resolution Statusopen · 3 weeks

Each colour links a phrase in the raw text to the Dimensions extracted from it. Hover either side.

Sentiment is a question

You probably know sentiment analysis, which sorts text into positive, negative and neutral. Think of it as one simple research question asked of every conversation:

Question: What is the tone of the conversation?
Answer: positive, negative or neutral

Asked across thousands of conversations, the answers add up to a share, which is how sentiment is usually reported. For example, a finding could read “X% of support conversations in March had a negative tone.”

Dimensions answer new questions

A Dimension works the same way, but the question can be anything the text can answer. Because a language model reads the conversation, you classify with a prompt instead of training a machine learning model.

Question: What was the outcome of the conversation?
Answer: complete, escalated or abandoned

Across thousands of support conversations, the share of escalations becomes clear. For example: “X% of live chat support conversations in June were escalated to a manager.”

Sentiment tells you the tone of an interaction. Dimensions can also tell you what happened, why it happened, and what to do about it.

Kinds of questions

Kind	Example Dimension	The question it asks	What it’s good for
Annotation	Products	Which products were mentioned?	Finding the products, services or features behind the most contacts
Scoring	Predicted Rating	How would the customer rate this interaction from 1 to 10?	A satisfaction signal on every conversation, without a survey. See Predicted scores.
Summarization	Predicted Feedback	How would the customer describe this conversation?	Seeing the service from the customer’s side, and the gap between your metrics and their experience
Recommendation	Predicted Solution	What could be done to resolve this issue next time?	A running list of fixes drawn from real conversations

Open-ended questions and dynamic labels

Keyword searches, trained models and fixed segmentations all decide the categories in advance. A Dimension doesn’t have to. It can generate a label from what the conversation says, so themes you didn’t plan for still show up.

Reason: what is the reason the customer reached out to support?
Category: what is the overall theme of the conversation?

Labels written this way are called dynamic labels, for example order tracking inquiry or refund request.

Clustering answers into levels

Some Dimensions come in pairs: a detailed answer and a cluster that groups similar answers. Reason and Category work like this, and so do Predicted Solution and Solution Cluster:

Predicted Solution: what is the potential resolution to the customer’s issue?
Solution Cluster: what are the themes across those resolutions?

The pair gives you two levels of analysis: the cluster to count, the detailed answer to read. You can also ask the agent to group dynamic labels into themes for you.

How a Dimension gets made

A prompt defines a set of Dimensions: a name, instructions, and for each Dimension its type and allowed values.
Enrichment reads each conversation in full (all the messages that share a sessionId; see Sessions and conversations) and fills in every Dimension of the prompt.
The results land in a table named after the prompt, one row per conversation, one column per Dimension.

Prompt support_reasons

reasoncategorybilling, delivery, product_quality, none_detected
competitorfree texta name, or none_detected
effortinteger1 to 5
refund_askedbooleantrue, false

Table: support_reasons

sessionreasoncompetitoreffortrefund_asked conv-1001deliverynone_detected4true conv-1002billingnone_detected2false conv-1003product_qualityAcme Beans3false

One row per session, one column per Dimension. Names and values are examples.

Dimensions are read from the conversation’s text only. The model doesn’t see your extra columns (rating, region, plan). Those are already structured, so you join them in afterwards.

For the ideas behind this, see How enrichment works.

Standard and custom Dimensions

There are two kinds, and the platform uses its own names for each:

	Standard Dimensions	Custom Dimensions
Called in the platform	System Prompts	Custom Prompts, or Prompts
Where to find them	The Prompts tab	The Prompts tab, once saved
Who defines them	Dimension Labs	You, or Dimension Labs with you
Covers	General questions that work on most text: Reason, Category, Predicted Rating, Summary and more	The signals specific to your business, in your own terms
When they run	Every day only if Dimension Labs turns on daily enrichment for your workspace; otherwise by backfill	Every day on live integrations they’re applied to; by backfill on uploads and history
Level of analysis	Session and section	Session
Page	Standard Dimensions	Design custom Dimensions

Start with the standard set for a baseline, then add custom prompts to go deeper where it matters.

Output types

Every Dimension has one of five output types. The same names are used on every page; the JSON for each is on the config reference.

Output type	Use for	Example values
Fixed category	Classifying into a list you define	`billing`, `delivery`, `product_quality`, `none_detected`
Dynamic label	Short labels the model writes from the text, with no list	`order tracking inquiry`
Verbatim quote	The customer’s exact words on a topic	`"the app logged me out twice"`
Score	A whole number on a defined scale	`1` to `5`
Boolean	Present or not	`true`, `false`

Decimal numbers aren’t supported. Use a whole-number scale instead.

none_detected and empty values

none_detected means the model read the conversation and found nothing for that Dimension. It’s an answer: “this customer didn’t mention a competitor”.
An empty value means nothing came back for that conversation. For a score, an empty value is the right answer when there isn’t enough signal: see Predicted scores.

Count them separately. A Dimension that is mostly none_detected isn’t broken: a signal that appears in 5% of conversations can be the most important one you have.

Reading Dimension values

Everything is stored as text, including scores and booleans. Ask the agent to treat a score as a number when you want an average.
The model can occasionally return a value outside the list you defined, or a spelling variant (none detected with a space). The agent can group these, and a cluster of unexpected values is sometimes a real signal worth a look.