Conversations

Conversations display the AI interactions across your organization, providing visibility into the User prompts and AI responses generated when working with AI Apps.

Conversations enable auditing, reviewing conversations, identifying risky or harmful activity, and help ensure compliance with organizational policies as well as government regulations.

WitnessAI supports multi-language interactions internally so your policies and guardrails can evaluate your AI data and protect your AI services as intended. Non-English text is not currently displayed as English in the Console.

Step-by-Step Instructions

Filtering Conversations

The Conversations table, with the date range, Group, User, Application and Show Beta Apps filters above it, and Search Conversations, Download and the column picker on the right.

Access the Conversations page from the left-sidebar menu.

You will see a list of recent conversations between Users and AI Applications, newest first. The summary above the table counts the conversations that match, for example 138 Conversations in last 7 days, and updates when you change a filter.

Use the Date range picker to choose your time window, and the Group, User and Application filters to narrow the list. The date range starts at Last 24 hours. On larger deployments the picker limits how long a window you can choose: 31 days on large deployments and 7 days on extra-large ones. Group, User and Application each accept several values, and the values you've chosen are listed first.

A filter applies as soon as you choose a value. There is no Apply button. Select Clear to reset all filters.

To change the order, select the Date & Time column heading. Each click switches between oldest first and newest first. The other columns can't be sorted or filtered from their headings; use the filters above the table instead.

Exporting Conversations

Select Download above the table, then Download CSV or Download JSON. If you've checked any rows, only those rows are exported. Otherwise every row on the current page is exported, up to 100 by default. To export a larger set, narrow your filters to the conversations you need and export page by page.

Viewing File Attachments in Conversations

Enter “upload” in the Search Conversations field. Conversations that include files will display. Clicking on a row will display the slide-out with full details.

Conversation Details for a file upload, with the processed file named in the prompt and the Blocked and Redacted badges.

The Conversation Details slide-out sidebar displays information on files processed. Text files less than 100MB are supported. Larger files and files containing malware are blocked.

Files contained in AI App responses are not scanned or blocked, allowing users to save them as needed.

Reviewing Conversations

Consider a scenario: You are a team manager, and you want to review how your team uses AI. Pick a Date Range and a Group, then have a look at a few Conversations.

Click on a row to view the details. The slide-out panel will display the full conversation.

The Conversation Details panel showing a user prompt, the AI application's full response and the detected intention.

The slide-out panel display

Initial Prompt, Latest Intention, User Group, User Email, AI Application, Number of Prompts, Number of Alerts, and Date & Time.
Step through every user prompt and AI response in the Conversation. Use the left and right arrows to navigate between prompts, or click to jump to the beginning or end of the conversation at any time. The scroll bar on the right side moves automatically as you page through prompts.
View the complete Prompt submitted by User, with the timestamp and intention.
View the complete Response from the AI App.

Viewing Alerts in Conversations

Click on a Conversation Row displaying text in the “Risk Type” column.

Conversation Details for a blocked prompt, with the Blocked and Redacted badges, View Alert, and the expanded details: topic, subtopic, policy, prompt ID and latency values.

Alert Indicators: Alerts are now more prominently displayed and viewable within the slide-out panel when you click on a conversation row. Including the topic, subtopic, policy, and prompt. Click "Show Details" to expand the alert information, or "Hide Details" to collapse it.
Sensitive Data Redaction Display: Redacted content is clearly marked in the prompt. If you have the appropriate role, you can click on the redactions to see the original text. Use the “Sensitive Content” slider to show or hide sensitive content.
Guardrail Actions: View the actions that occurred on the right side, including whether content was redacted, if a warning was shown to the user, and if the prompt was blocked or allowed to be sent to the LLM.
Navigate to Alerts Console: Clicking "View Alert" navigates to the Alerts console, where you can see full details of the specific alert including all Guardrail actions triggered, the policy and its version, and which guardrails were triggered.

Viewing Alerts Details

The Alert Details panel on the Overview tab for a blocked file upload, with the Policy, Guardrails, Conversation, Prompt, Attachments and Classification sections.

Click on the eye icon to mask or unmask sensitive data
View classification, source and destination, prompt ID, and conversation ID
Click on "Matches" to expand details showing specifically what was matched to the guardrail
Add labels to alerts as desired. Up to three existing and custom labels can be included in any alert.
Scroll down to see more details

Viewing Latency Measurements

Conversations now include Latency Measurements for most prompts and responses. They are also called Latency Metrics.

Click any conversation. The Conversation Details panel slides out from the right.

Hover over the prompt, then click Show Details.

The Conversation Details panel with the Show Details link below the prompt's intention and action.

The Latency Measurements will be displayed at the bottom of the column.

The expanded details column: topic, subtopic, policy and prompt ID, then Input Evaluation, Output Evaluation, Total Evaluation, Provider Response Time, Request Payload Size and Response Payload Size.

Clicking the “Hide Details” button will collapse the details column.

Understanding Latency Measurements

These metrics are captured for each round-trip LLM prompt → response lifecycle. Actual naming varies by vendor or internal instrumentation.

Here’s how they’re typically interpreted and how they relate to each other.

Input Evaluation

What it measures

Time spent processing the prompt before response generation (inference) starts

Usually includes

Request parsing & validation
Tokenization of the input prompt
Safety / policy checks
Prompt templating or system-message injection
Routing to a model instance (queueing, load balancing)

Output Evaluation

What it measures

Time spent generating and evaluating the model’s output

Usually includes

Token-by-token generation
Decoding
Output safety checks
Post-processing (formatting, truncation, JSON validation)

Provider Response Time

What it measures

Time the AI provider takes to return its response

Request and Response Payload Size

What they measure

The size of the request sent to the provider and of the response it returns

Total Evaluation

What it measures

End-to-end request latency (client → model → client)

Conversations Console Whats New in Conversations?Latency Measurements Step-by-Step Instructions Filtering Conversations Exporting Conversations Viewing File Attachments in Conversations Reviewing Conversations The slide-out panel display Viewing Alerts in Conversations Viewing Alerts Details Viewing Latency Measurements Understanding Latency Measurements Input Evaluation Output Evaluation Provider Response Time Request and Response Payload Size Total Evaluation