AI, socket and PII logs

The AI log records every AI request the platform makes: the calls behind Mind chats, embeddings, model listings and tests. Use it to investigate a slow or failed answer, see which model a request used, or read the exact request and response that went out. It sits in the Logs tab of Observability, next to Audit, SSO Logs and Notifications, which are covered in Audit, sign-in and notification logs.

Open the AI log

Open Observability from the sidebar, then click Logs.
In the row of sub-tabs, click AI.

A sub-tab appears only if your role gives you its permission. See Permissions below.

Investigate an AI request

The AI table lists each AI request with Date & Time, Type, Path, Model, User, Status, Duration, Aborted, Error, Agent Type, Conversation ID, Round ID, Request Body and Response Body. Further columns include Finished At, Request ID, User Identifier and User ID.

Open Observability > Logs > AI.
Set Period to Today, Last 7d, Last 30d or Custom.
Use the column menus on User, Model or Conversation ID to narrow the table. Filter option lists load from the server and support search.
To find slow requests, sort by Duration. Hover a duration to see the exact value in milliseconds.
Click a row to open its details in a resizable panel on the right. The selected row stays highlighted; click it again to close the panel.

The panel header shows the type, status, duration and speed label, with an Aborted marker when the request was cut short. The line under it shows the user and the time range. Below that it lists Request ID, Conversation, Round, User Identifier and User ID, each with a copy button, then any error, then the Request body and Response body sections. An empty body reads No request body or No response body; an error without text reads No message, and an aborted request adds a Request was aborted note.

The speed label depends on the duration:

Label	Duration
Fast	Under 1 second
OK	From 1 second to under 10 seconds
Slow	10 seconds or more

Type names the kind of call, for example chat, models, embedding or test.
Status shows the HTTP status, coloured by family: 4xx amber, 5xx red.
Aborted marks a request that was cut short; Completed marks one that finished.
Identifier and body cells copy when you click them; a check mark confirms the copy. A value that is too long to fit shows in full on hover.

Records marked Content not logged

If Request Body or Response Body shows Content not logged, the user has switched off DataMind Oda content logging. Only token usage and costs were recorded for that request, so the bodies are not available here. Hover over the badge to read the reason.

Export the AI log

The AI log exports as JSON only. Excel cells cannot hold request and response bodies of this length, and CSV and TSV flatten their nested content. An export holds at most 300,000 rows. Exporting needs only View AI Logs, and the file is named after the log and the date, for example ai_gateway_logs_2026-10-16.json. The export contains the rows that match your period, filters and search.

Related logs

Socket logs follow a chat's live messages and explain why one failed.
PII detection logs show each decision PII detection made about typed text, and the detections behind it.

Permissions

To do this	You need
Open the Logs tab	View Audit Log
See the AI sub-tab	View AI Logs, preset AI Log Viewer
Export the AI log	View AI Logs

Note

The Observability Administrator and Observability Viewer presets both include AI Log Viewer, so both can open the AI sub-tab. The PII sub-tab has its own permissions; see PII detection logs.

Periods and limits

The AI log's Period offers Today, Last 7d, Last 30d and Custom. It starts on Last 7d.
A custom range can span at most one year. A longer range shows Date range too large, with the message Custom date range cannot exceed 1 year.
Your browser remembers the last period you chose and your column layout. Neither is stored on your account.
Rows per page starts at 10; you can pick 10, 20 or 50.
The details panel has copy buttons for IDs and bodies, so you can paste them into another tool.

Note

If a refresh fails after the table has loaded, the table keeps its rows and shows an error bar with Retry instead of clearing your search. If it keeps failing, see Observability troubleshooting.

Questions, answered

A Mind answer took very long. How do I check?

In AI, set Period, filter User to the person who asked, and sort by Duration. Open the record. Its speed label reads Slow when the request took 10 seconds or more.

Why are the request and response empty for some AI records?

Those records show Content not logged. The user switched off DataMind Oda content logging, so only token usage and costs were recorded.

Why can I export the AI log only as JSON?

Request and response bodies are too long for Excel cells, and CSV and TSV flatten their nested content. JSON keeps the structure.

How do I find every AI request from one chat?

Copy the Conversation ID from any row of the chat and filter the Conversation ID column to it. Every request from that chat shares the id, so you can read them together.

What does Aborted mean in the AI log?

The request was cut short before it finished, for example because the user stopped the turn. The panel adds a Request was aborted note. Filter the Aborted column for Aborted or Completed.

Does the AI log remember my column setup?

Yes. Your column layout, and the period you chose, are remembered in this browser. They are not stored on your account.

Which permission shows the AI tab?

View AI Logs shows the AI sub-tab. You also need View Audit Log to open the Logs tab at all.