Conversation History and Trace Details

Every time a user chats with a Digital Worker, FPT AI EOS Console records the whole conversation. The Insights tab gives you two levels of detail:

History: The list of conversations, replayed exactly as the user saw them.
Trace details: A turn-by-turn breakdown - Which tools the Digital Worker called, whether a Guardrail stepped in, how long each step took and how many tokens it used.

Open Trace details when you need to answer questions like: Why was this answer wrong? Why did the user wait so long? How many tokens did this conversation use? Did the Guardrail block the right thing?

Open conversation History

Open the Digital Worker

In Console, select Digital workers in the left menu and open the one you want, for example Hỏi đáp NQ_57.

Go to Insights

Click the Insights tab at the top, then choose History in the left column.

Pick a conversation

Click a row in the table. The conversation panel opens on the right and shows every message in the order the user saw it.

Columns in the History table

Column	Meaning
End time	When the last message of the conversation was sent
Conversation ID	Unique identifier, used for search or when contacting support
Channel	Where the conversation happened: Workspace, Live Chat, API, Zalo…
User	Name and email of the person chatting
Triggered by	How the conversation started, for example a user chatting directly
Message count (last column)	Total number of messages in the conversation

Search and filter

Search box: Search by content, sender name, conversation ID or message ID.
All channels: Show conversations from one channel only.
All time: Limit to a date range.

Actions in the conversation panel

Button	What it does
Link icon	Copy conversation link so a teammate can open this exact conversation
View trace details	Opens the Trace details page for the conversation
Copy icon next to the ID	Copies the conversation ID
X	Closes the panel

Open Trace details

In the conversation panel, click View trace details. The page has three columns:

Column	Content
Turns (left)	Every question-and-answer turn, each with its latency and token count
Turn details (middle)	The user’s input and every step the Digital Worker took to answer
Stats of conversation (right)	Time, latency and token totals for the whole conversation

Click Back to History at the top to return to the list. The copy icon next to the ID in the header copies the conversation ID.

Turn list

A turn is one message from the user plus everything the Digital Worker did to answer it. Click a turn to jump to its details in the middle column. Each turn card shows:

The turn number and the start of the user’s question.
Latency: Time from receiving the question to finishing the answer, for example 27.4s.
Tokens used in the turn, for example 109.8K (K = thousand, M = million).
Wrench icon: The turn called a tool.
Shield icon: A Guardrail blocked something in this turn.
Orange dot and a red label (such as “Input guardrail blocked”): The turn has an Issue worth a closer look.

Hover over a turn card to see that turn’s own numbers: Start and end time, latency, time to first token and token breakdown.

Metric	Meaning
Start, End	When the Digital Worker received the question and when it finished answering
Latency	Time from Start to End - How long the user waited for this turn
First token	Time from receiving the question until the first word of the answer appeared
Token breakdown	Input, Cache read, Output, Reasoning and Total for this turn only. See Stats of conversation below for what each item means

Turn details

Each turn has two blocks:

Inputs: What the user sent, labelled Guardrail passed when it broke no rule.
Outputs: Every message the Digital Worker produced in the turn, in order. The top-right corner shows the Number of messages.

Each Digital Worker message (labelled AGENT) has an info line on the right, for example deepseek-v4-flash · 6.0s · 21.5K tokens · Guardrail passed:

Item	Meaning
Model name	The model that produced this message
Time	How long the model took to produce the message
Tokens	Tokens used by this model call alone: Input plus Output
Guardrail passed	The message was checked by Guardrails and broke no rule

A turn can contain several AGENT messages, because the Digital Worker may call the model more than once: Call a tool, read the result, then answer. Click the arrow next to TURN to collapse or expand the whole turn.

Inspect a tool call

Tool calls appear under the message, for example vbpl-knowledge__knowledge_search - The Digital Worker searching the knowledge base through the vbpl-knowledge Connector. Click a row to expand it:

Field	Meaning
Connector	The connection the tool belongs to
Tool call id	ID of this call, useful when matching logs
Status	Whether the call succeeded or failed
Input	Parameters the Digital Worker sent, such as the query and filters
Output	What the tool returned
Duration (right)	How long the call took, for example `1.6s`

The Input and Output blocks can be shown as JSON (formatted) or Raw, and each has a copy button.

Wrong answers often start at the lookup step. Open Input to see which query the Digital Worker searched with, then open Output to see what the knowledge base returned. If the results look right but the answer is still wrong, review the Instructions.

When a Guardrail steps in

A turn where a Guardrail intervened is labelled Issue. In Outputs, the Guardrail · input check row shows its result (for example Refused). Click the row to see:

Rule: The rule that fired, for example __injection__.
Screen: The type of check, for example injection (a prompt trying to override the Digital Worker’s instructions).
Action: What the Guardrail did, for example Refused.

Example: A user tells the Digital Worker to ignore its instructions and make up figures. The Guardrail refuses, the message never reaches the model, and the user gets the safe reply shown in the AGENT message below it.

See how to set up rules in Guardrails.

Stats of conversation

The right column sums up the whole conversation.

Group	Metric	Meaning
Stats of conversation	Turns	How many times the user asked and the Digital Worker answered
Time	First start, Last end	When the first turn started and the last turn ended
Latency	P50	Typical wait: Half of all turns finished faster than this
	P99	Worst-case wait: 99% of turns finished within this
	First token P50	Typical wait before the user sees the first word appear
Token breakdown	Input	Tokens the Digital Worker reads before answering: Instructions, chat history, tool results
	Cache read	Input reused from cache, cheaper than reading it fresh
	Output	Tokens the Digital Worker writes: The answer and tool calls
	Reasoning	Tokens the model spends “thinking” before answering, if it supports this mode
	Guardrail checks	Tokens used to check questions and answers against Guardrails
	Total	Sum of the above

K means thousand and M means million. The percentage in brackets is each item’s share of total tokens.

Understanding P50 and P99

Line up every turn from fastest to slowest:

P50 is the turn in the middle - What most users experience.
P99 is the turn near the very end - The longest waits.

Example: A P50 of 15 seconds and a P99 of 65 seconds means users usually wait about 15 seconds, but at worst more than a minute. The further apart the two numbers, the less predictable the response time.

Use the stats to improve your Digital Worker

Input takes almost all tokens

The Digital Worker re-reads the whole chat history every turn, so long conversations get more expensive per turn. Consider shortening the Instructions and limiting how much the knowledge base returns.

Cache read is zero

Content repeated across turns is billed at full price. Choosing a model that supports caching can cut costs significantly.

P99 is many times P50

Open the slowest turns in the turn list to see whether the time goes into the model call or a tool.

Many turns with Issues

Check which Guardrail rule is firing. If it blocks valid questions, adjust the rule’s description and examples in Guardrails.

Create, edit, delete

History and Trace details are read-only. Console records each conversation automatically, so you cannot create, edit or delete conversations here. To change how the Digital Worker answers, update its configuration in the Build tab and Publish again.

Performance in the Insights tab is under development and coming soon.

Introduction

FPT AI Console

Conversation History and Trace Details