What the agent can see

The agent’s limit is not what it can think about. The limit is what data it can see. The range of questions it can answer well grows with the data in your workspace: each new source and each new Dimension makes it more useful.

The agent reads the tables below. Table names in your workspace depend on what you’ve uploaded and what your prompts are called, so the agent always lists them first. To use your Dimensions outside the agent, pull them with the Export API.

Table	What’s in it	One row is
transcripts	Every message: text, direction (incoming), time, sessionId, and intent where your data has one	A message
structured_metadata	The extra columns you uploaded with your conversations, such as rating, store, plan, channel	A conversation
One table per prompt, named after it	The Dimensions that prompt extracts	An enriched conversation
One table per business data file, named after the file	Whatever you uploaded	Whatever its grain is

A new workspace may have only transcripts and structured_metadata until a backfill has run.

Three kinds of fields

Kind	Where it comes from	Where it lives
Metadata	The columns you upload with your conversations, and your business data files	structured_metadata, and one table per business data file
Standard Dimensions (System Prompts in the app)	General fields every workspace has available, such as sentiment and contact reason	system_session_prompt
Custom Dimensions	Fields you design for your business	One table per prompt, named after it

Standard Dimensions are only filled in where they’ve been run: by a backfill from the Backfill tab, or every day if Dimension Labs has turned on daily enrichment for your workspace (ask [email protected]). See When enrichment runs. To get more out of your data, add Dimensions for the questions you care about. See Design custom Dimensions.

Ask it to show you

• What’s in my workspace?
• List my prompts.
• List my Dimensions.
• What metadata do I have?
• Show me the tables in my workspace and how many rows each has.
• What columns are in structured_metadata?
• Which Dimensions does my reviews prompt produce, and how often is each one filled in?

The agent lists what’s there and tells you which kinds of questions it can answer well with it.

Find where something is captured

When you don’t know whether, or where, your data holds something, ask the agent to look:

• Do we track competitor mentions anywhere?
• Where do we capture cancellation reasons?
• What Dimensions do I have for sentiment?
• Help me find the right Dimension for measuring customer effort.

Then drill down into what it finds:

• Show me ten conversations where that Dimension says “price”.
• How is that field filled in across channels?
• What date range does it cover?

If nothing captures the concept yet, the agent says so and can help you draft a new Dimension. See Designing Dimensions with the agent.

Things to know when you read its numbers

• A prompt’s table can hold fewer rows than transcripts holds conversations: only enriched conversations appear.
• Counting rows in transcripts counts messages, not conversations.
• Everything uploaded is stored as text. The agent converts numbers and dates when it calculates.
• The platform’s date fields are in US Pacific time, so a record near midnight UTC can fall on the previous day. The agent’s own clock runs on UTC, so give explicit dates for “last month” or “this week”.

See Numbers look wrong if counts don’t match what you expect.