> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://api.logicdialog.ai/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://api.logicdialog.ai/_mcp/server.

# Data collection and topics

Use **Evaluation** to configure both conversation success criteria and information to collect. Make changes in a workstream, save them, then deploy to the required environment. Saving alone does not change running assistants.

## Configure topics

Open **Evaluation → Data collection → Topic**. Topic is a built-in text field with the reserved identifier `topic`. Its name, type and extraction instructions cannot be changed, and it cannot be deleted.

Add each possible topic value and a short description of when it applies. Values are exact, case-sensitive identifiers. Keep them stable if routing or reporting depends on them. For example:

| Value      | When to use it                                |
| ---------- | --------------------------------------------- |
| `billing`  | Charges, invoices, payments and refunds       |
| `delivery` | Tracking, missing parcels and delivery delays |
| `account`  | Login, password and account access problems   |

`unknown` is always included. It covers unclear requests, topics outside the configured choices, and multiple unresolved issues without an explicit priority. Without configured topic values, all classifications use `unknown`.

The platform generates the classification instructions from this definition. During live human handover, the assistant supplies a topic with its request. LogicDialog validates it against the deployed configuration and uses `unknown` for missing or invalid values. Duplicate requests cannot change the topic of an accepted handover. Classification remains probabilistic: clear, distinct descriptions and representative testing are important.

For CXone, configure **Channels → CXone → CXone contact topic field identifier** to forward this value. Configure the corresponding routing rules and an `unknown` fallback in CXone. Live handover must be enabled in **Channels → Agent platform**.

## Add other data fields

Choose **Add data field**. Enter a unique identifier, display name, type and extraction instructions. Supported types are text, yes/no, whole number and number. Text fields may optionally restrict the allowed values. `topic` is reserved; other identifiers must start with a lowercase letter and contain only lowercase letters, numbers or underscores.

Custom fields are collected after the AI conversation ends. They are not available as live handover arguments. Removing a field stops future collection after deployment; existing results remain available.

## Review collected results

Open a conversation's **Collected data** tab to see extracted values and the reasons recorded during analysis. False and zero are displayed as collected values, rather than treated as missing. Results may not be available immediately when a conversation ends.

The **topic at handover** is stored separately from the final analysed topic. The final topic may differ when the conversation contains a correction or a change in priority. Final analysis never changes an earlier routing decision. In a human handover, the AI conversation ends when the human relay begins; subsequent human messages are not part of that analysis.

## Explore topics in Performance

Open **Performance → Topics** and select the environment and date period. Counts use the final analysed topic where available, otherwise the saved handover topic. Conversations without either are excluded from topic counts; `unknown` is a real fallback category, not missing data.

Select a topic to open **Conversations** with the same environment, start date and exact topic filter. Topic counts use the loaded conversation sample; the page indicates when more history exists. Historical values remain visible even if their definitions are later renamed or removed.