Overview

A Transcript is the saved record of a conversation with your agent. Open Transcripts in the project sidebar to find a session, read its messages and inspect the execution logs, metadata and available Evaluation results.

Find and review conversations in the Transcripts table. Photo:

Examples of when to use Transcripts

Use Transcripts when you need the detail behind a customer interaction. You can:

  • Investigate an unexpected answer: read the customer’s request, the agent’s reply and the logs for that turn.
  • Review a release: filter to an Environment and inspect conversations after a change.
  • Analyze a segment: attach properties such as a customer’s plan tier or inquiry topic, then filter by them.
  • Score historical conversations: run an Evaluation against an existing set of Transcripts.

Understanding the Transcripts table

Each row represents a conversation. Conversation fields can include its date, Environment, user ID, cost and duration, with other available metadata, custom properties and Evaluation results. A missing value means it is not available for that Transcript; do not treat it as a zero or a successful score.

Use Analytics for aggregate trends and Evaluations to define how conversations are scored.

How to use Transcripts

1

Find the conversation

Open Transcripts and use Date & time, Environment or another relevant filter to narrow the list.

Find and review conversations in the Transcripts table. Photo:
2

Open a Transcript

Click a conversation row. Compare its metadata with the messages to confirm you have the session you intended to review.

Review the conversation and its logs together to trace what happened. Photo:
3

Trace the relevant turn

Read the conversation alongside Logs. Check the input, output and any relevant execution details around the unexpected behavior, then use the evidence to decide what needs changing.

Review the conversation and its logs together to trace what happened. Photo:

After changing the agent, test the same situation again and review the new conversation. Keep the original Transcript as evidence of the behavior you were investigating.

Finding conversations

Narrow the list to the sessions relevant to your investigation.

Use the filter chips above the table to choose a date range, Environment, user ID, cost, duration, Evaluation result or other available metadata. Custom properties can add your own segmentation criteria. Combine filters to focus a review, then clear them when returning to the full list.

Find and review conversations in the Transcripts table. Photo:

Check the active chips when a conversation appears to be missing. In particular, confirm its date and Environment are included before concluding that it was not saved.

Review a release’s conversations

Choose a consistent sample for checking a recent agent change.

For a release review, select the Environment that received the change and a date range after it was published. Read several relevant conversations, then compare with a period before the release using the same scope.

If you add an Evaluation-result filter, remember that you are reviewing scored conversations that match that result. That subset may not represent every conversation in the period.

Reviewing a conversation

Read messages, metadata and execution logs together.

Click a Transcript row to open the detailed view. The conversation appears beside its metadata, available Evaluation results and Logs. Confirm the user, date, Environment and version before investigating an individual turn.

Review the conversation and its logs together to trace what happened. Photo:

Read an Evaluation’s score and reasoning against the actual exchange rather than treating the score alone as a diagnosis. Where conversation history is available, use it to understand earlier interactions with that user. Voice conversations can include a recording when one was captured; availability depends on the session.

Logs provides a timestamped view of inputs, outputs and available execution details, warnings, debug information and credit usage. Use the entries around the relevant turn to check which resources ran and what they returned. A missing value is not evidence that an action succeeded.

Conversation review best practices

Connect an observed symptom to evidence before changing the agent.
  • Start with the customer’s goal. Identify what they asked and what a useful result would have been before assessing the reply.
  • Follow the sequence. Compare the input, resource execution and final response at the same turn; a later outcome may not explain the earlier decision.
  • Check the recorded version. Review which version handled the session so a fix to the current draft is not confused with historical behavior.
  • Verify a change with a new run. An old Transcript documents the original interaction; it does not change when you edit the agent.

Transcript properties

Store metadata for later filtering and analysis.

Properties attach metadata to a Transcript for later review. Use them for fields such as a customer’s plan tier or the topic of a conversation. They differ from Variables, which the agent uses while the conversation runs.

In a Workflow’s Set step, add the field under Properties to set and provide its value. Use Manage properties in Transcripts to manage the available property names.

Find and review conversations in the Transcripts table. Photo:

Run the relevant path and inspect the resulting Transcript to confirm the property was recorded with the intended value before using it as a reporting filter.

Transcript properties best practices

Keep segment names and values consistent across conversations.
  • Choose stable names. Use one field for a concept such as plan tier instead of several near-duplicates.
  • Normalize categorical values. Consistent values make filters useful; differently spelled values can split the same segment.
  • Record at the relevant point. Set the property when the Workflow has the information, then verify the saved value on a Transcript.

Table columns

Display the metadata and scores your review needs.

Click the Edit table view icon in the Transcripts toolbar. Choose the default fields, available Evaluation results and custom properties to display, and arrange them in the order you want to review. Click Save to apply the view or Reset to default to restore it.

Choose which conversation fields and Evaluation results appear as columns. Photo:

Adding a column changes the view; it does not populate missing metadata or run an Evaluation. Open a row when you need the underlying conversation or the explanation behind a score.

Saving Transcripts

Choose whether tests and sessions without a reply are recorded.

Open the Transcript settings icon in the Transcripts toolbar. Save tests to transcripts keeps internal test conversations available for review. Save transcripts with no interaction saves sessions where the user never responded. Set each option for your review needs, then click Save.

Choose whether internal tests and sessions without a reply are saved. Photo:

Saving tests can support QA investigations, while omitting sessions with no interaction can reduce noise. Check these preferences when investigating missing test sessions or unexpectedly quiet conversations. Review the Environment and date filters as well; recording preferences and the current table view are separate controls.

Batch Evaluations

Score historical conversations with selected criteria.

Open Batch run evaluation from Transcripts. Choose the Evaluations to apply and use the dialog’s filters to define the historical conversations to include. Check the criteria, scope and expected credit use before clicking Run.

Choose Evaluations and filters before running them on historical Transcripts. Photo:

This lets you backfill an Evaluation created after those conversations, or reassess a historical sample after refining criteria. When processing finishes, inspect individual results as well as aggregate scores. Batch processing consumes Evaluation credits; it is separate from enabling an Evaluation for future Transcripts.

For metric types, criteria and result interpretation, see Evaluations.

Use the up and down arrow keys to select a result, Enter to open it, and Escape to close the search.