Analytics datasets and intent hierarchies¶
Last reviewed: 2026-10-02
Use Data → Manage dataset to provide the inputs an Analytics Agent needs. A dataset contains multiple tables, not just transcript files.
Supply the data for your questions¶
| Input | What it enables |
|---|---|
| Transcripts or audio | Conversation inspection and evidence search |
| Contact metrics | Handling time, observed transfers, containment and operational dimensions |
| Surveys | CSAT and NPS analysis linked to contacts |
| Customers, agents and queues | Segment and organizational breakdowns |
| Operational changes | Compare policy, process or staffing changes with contact trends |
| Flight disruptions | Relate flight events to airline contact patterns |
| Reservations and proactive messages | Analyze downstream outcomes and outbound interventions |
| Custom tables | Additional structured business facts with your own columns and relationships |
Preparation generates annotations, embeddings and prepared insights when those capabilities are enabled. It does not invent operational records from transcript wording. For example, escalation language does not establish an observed transfer, and a conversation's duration is not automatically its handling time.
Import tables¶
- Upload and prepare transcripts in Files & conversations.
- Open Tables & inputs, then Import tables.
- Choose a table type or Custom table. Download its CSV or JSON template.
- Select a file and map its columns. CSV uses one business record per row. JSON accepts row objects, an exported table, or a complete table bundle.
- Review the preview, unique row key, relationships and import behavior.
- Choose Save inputs, or Save inputs & prepare for chat.
Replace changes this table's records while keeping other tables. Append adds records and rejects duplicate keys. Update matching keys updates existing records and adds new ones; unmapped columns keep their previous values. Explicitly mapped empty cells mean unknown.
Use the same conversation IDs as your transcripts. Supply ISO 8601 dates, handling time in seconds, and disruption delays in minutes. NPS input is a respondent's 0–10 score; CSAT is 1–5. Boolean fields accept true/false, yes/no or 1/0. Array/object fields use JSON.
Import related tables together in a bundle, or in stages. Unmatched IDs in a populated related table stop the import. Missing reference tables produce warnings identifying what to add. The built-in table types retain their required keys and relationships.
Table imports require builder or administrator access. Limits are 100 MiB per request, 200,000 rows per table and 100 columns. Custom identifiers use lowercase letters, digits and underscores and cannot begin with a digit.
Explore and verify¶
The table list shows readable names, row counts and whether data was supplied or generated. Select a table to:
- Search all its rows and browse 50-row pages.
- Inspect columns, types, unique keys and relationships.
- See the imported filename and table description.
- Export CSV or JSON, or edit/replace a supplied table.
Table version switches between the latest supplied inputs and a completed prepared snapshot. Export bundle downloads the selected version's typed tables and available report bodies. Reimporting it through Tables & inputs imports its supplied tables; preparation regenerates derived tables. Legacy full prepared-bundle import remains under Prepared insights for migration comparisons that must retain original annotations and reports.
Coverage counts show how many conversations have dates, annotations, semantic indexing and handling-time records. Unknown values remain unknown. A warning appears when the supplied inputs do not have a prepared result for the selected transcript version.
Build and select the chat version¶
Open Prepared insights to follow the build and select its completed result for new chats. The automatic source option includes the latest supplied inputs. You can explicitly choose a previous prepared source or a transcript-only build.
When adding transcripts, preparation carries existing operational inputs forward. New conversations have unknown operational metrics until you supply them. Rows linked to removed conversations are excluded from the new chat version; earlier input versions and existing chats remain intact.
Saving inputs alone does not change an existing chat. A pinned agent keeps its selected prepared snapshot until you change that selection and publish.
Edit large intent hierarchies¶
Open Preparation → Annotations → Edit vocabulary & intent hierarchy. The wide modal includes drivers/root causes, intents, friction and escalation.
In Intent hierarchy:
- Expand or collapse branches, or use Expand all/Collapse all.
- Search labels, IDs, guidance or categories. Matching entries retain their parent path.
- Select an intent to edit its label, stable ID, guidance and parent.
- Select a category to rename it or move the entire branch.
- Add intents or create a category with its first intent.
- Import/export the intent list independently, or use the complete vocabulary import/export controls.
The editor supports 10,000 intents and ten category levels. Older category and subcategory data opens as a tree automatically. Arrow keys navigate the tree; Home/End reach its first/last visible entry.
Apply vocabulary saves the modal's changes into the agent's preparation settings. Cancel discards them. If another collaborator changes annotation settings while the modal is open, reopen it before applying your changes.
Stable IDs connect classifications across builds. Renaming labels preserves those IDs; changing an ID creates a different classification for future builds. Existing prepared results do not change until you rebuild. Large vocabularies also increase model input size, so choose a model with adequate context capacity.