For the complete documentation index, see llms.txt. Markdown versions of documentation pages are available by appending .md to page URLs.
Skip to content

Interaction Logs

Trace every step your workflows execute

The interaction log captures every internal step a workflow runs to produce an agent reply, from model calls to HTTP requests. Which steps run for any given reply depends on the workflow itself. The log is the raw record of what actually happened, and the starting point for any serious debugging or performance work.

You meet the interaction log in two places:

  • Observatory > Interaction Logs: Every interaction across every session, in one filterable table

  • Observatory > Sessions > Chat Sessions: Open any session and select a user message to see the trace behind it, scoped to that single conversation event

What Is an Interaction

An interaction is a single atomic step your workflow runs while producing a reply. Each type of step corresponds to a node in Designer that can fire during a workflow execution. Which types appear in any given trace depends on the workflow's design.

LLM

A call to a language model. Captures the model, prompt, conversation history, and the model's reply

Tool Call

A call to an MCP server or A2A agent. Captures the tool, the arguments sent, the output, and how the call ended

A semantic search against your Knowledge Base. Captures the (possibly rewritten) query and the matched documents

HTTP

An outbound HTTP request to an external system. Captures the URL, method, headers, request body, and the response

Variable

A write to a variable. Captures the variable name and the value the step set

Tags

A tag operation on the session. Captures the tag or tags the step applied

Redirect

A jump to another workflow. Captures the target workflow, including the variable template used to resolve it

Regardless of type, every interaction in the log carries the same surrounding metadata:

Field
What it captures

Timestamp

When the step ran, to the millisecond

Duration

How long the step took, in milliseconds

Errors

Whether the step succeeded or failed

Name

The friendly name of the Designer node that triggered the step, or the tool's name

Agent

The agent that owns the workflow

Deployment

The deployment that served the session

Session ID

A unique identifier of the parent conversation

Event ID

A unique identifier shared by all steps of a single user turn

Flow

The workflow the step ran in, even when the turn started in another one

JSON Payload

The raw request and response data, shaped to fit the step's type

Sessions, Events, and Interactions

Interactions sit at the bottom of a three-tier nesting: a session is one conversation, an event is one workflow execution within that session (typically one user turn), and an interaction is one step within an event.

Exploring the Interaction Logs

The Interaction Logs page at Observatory > Interaction Logs brings every interaction your agents have produced into one filterable table. Each row is one step a workflow ran, and selecting it opens the details in the side panel.

Filtering and Searching

Every Observatory page shares the same filter bar, which shapes the results you see. Work through them in order: scope the data, filter within it, then save the view or clear it to start over.

1

Set the Scope

Date Range and Agents define the data the page draws from. A wrong window or agent is the most common reason a page comes back empty.

Cover

Date Range

Set a start and end date to define the time window, then click Apply

Cover

Agents

Filter by agent or deployment, with your selected agents pinned first

2

Filter by Value

Within that scope, these filters narrow the results further. Each one lists only the values found in the current Date Range and Agents, so widen those if an option you expect is missing.

Filter
What it does

Tags

Include or exclude results by tag

Variables

Filter by a session variable's name, with an optional value to match against it

Contains

Keep sessions that include specific signals, such as user feedback or a CSAT response

3

Save or Reset Your View

Once the filters show what you want, keep that view or reset it.

  • Save: Stores your current selection (Date Range, Agents, and every value filter) as your personal saved view. Observatory reopens with it applied, and an indicator flags whenever your live selection differs from what you saved.

  • Clear: Resets the whole bar to its default view: the date range returns to its default and every other filter empties. Your saved view is left untouched, so Clear is a fresh start, not a delete.

Share a view: Filters live in the URL, so sharing or bookmarking the page reproduces the same view. Your selection also persists as you move across Observatory.

Use the search bar to find an interaction by the reference you have on hand:

Search by
When to use

Session ID

You already have a session reference from a link, screenshot, or bug report

Name

You know the node, tool or variable label (for example, csatSurveys) and want every step that touched it

Event ID

You have an event ID for a single workflow execution (typically one user turn) and want every step that ran in it

Variable rows are hidden by default. To show them, select Variable in the Type filter. Do this before you search for a variable name such as csatSurveys, or the search comes back empty.

Inspecting a Single Interaction

Select any interaction to open its details. This is where everything you have seen so far comes together: the interaction type, the event hierarchy, the raw payload, and more.

From any opened interaction you can:

Read the Payload

The raw request and response of the step in JSON, shaped to fit the interaction's type

Inspect Variables

A snapshot of session-scoped variables after this step ran, including any writes the interaction itself made

Reference the Step

The Session ID and Event ID for the step, each with a copy icon for use in tickets, reports, or the search bar on this page

View in Session

Jump back into the full conversation transcript with the user message that triggered this step already selected

View in Session

View in Session is the link from a single step back to the conversation that produced it. An interaction tells you what the workflow did, but not why it had to. The session view gives you that context: the user message that triggered the step, what came before, and what came after.

Use it whenever you have found a problematic step and need the conversation around it to judge whether the behavior makes sense. See Inside a Session for what you land on.

Logs in a Chat Session

When you want to debug one specific conversation, start in Observatory > Sessions > Chat Sessions. Select the session, then select a user message in the transcript to see the workflow steps that ran in response to it.

All the same interaction concepts apply here, just at a narrower scope: one user message at a time.

A few things to know about this view:

  • Each user message is the trigger that starts a workflow execution. Selecting a user message shows the steps that produced just that one reply.

  • Variable rows are hidden here too. Select Variable in the Type filter to show them.

  • Expand any step in the trace to see its Flow and its full details, including the raw payload

  • The Variables Values tab shows a snapshot of session-scoped variables at this point in the conversation, including the writes that happened up through this turn

To see a whole session's interactions at once, use the Interaction Logs page with a Session ID filter instead.

Working with the Logs

A handful of patterns cover most real debugging work. Each one starts from a different reference point and lands you on the relevant interaction in a few clicks.

Investigate One Session in a Single Table

When you already know the session you want to debug, the Interaction Logs page is often faster than the per-session view. Filtering by Session ID turns the table into the full chronological trace for that conversation. You see the whole trace at once, instead of one message at a time.

  1. Copy the session ID from the per-session URL or from the Session Details sidebar tab inside the conversation

  2. Open Observatory > Interaction Logs and switch the search mode to Search by Session ID

  3. Paste the ID and run the search

  4. Select Variable in the Type filter if you also need each variable write in the trace

  5. Sort by Timestamp, scan for errors or long durations, and select any row to inspect its payload

Find Slow Steps for a Specific Deployment

Performance investigations often start as "responses feel slow on Webchat lately" and end at a single LLM call or external request that crosses your latency budget. Combining the Agents filter with the Long duration preset narrows the table to the steps that matter.

  1. Set the Agents filter to the agent and deployment under investigation

  2. Open the Duration column filter and pick Long (>2s)

  3. Sort by Duration descending

  4. Open the slowest rows to see exactly what each one was doing

Triage Errors Across Agents

An error count spikes on a dashboard, or a teammate flags an issue without naming a session. The Errors filter surfaces every failing step in your time window. Pattern-spotting across that subset is usually how you find the root cause.

  1. Set the date range to cover the period of interest

  2. Open the Errors column filter and pick Yes

  3. Scan the table for patterns: a single agent, a single deployment, a single interaction type, or a single source node showing up repeatedly

  4. Open a matching row to read the error in the payload, then use View in Session to see the conversation around each failure

Find Every Step a Specific Node Produced

The Name column shows the friendly name of the Designer node that produced each step. Setting clear, descriptive names in Designer is what makes this search useful later: the Name search is only as good as the names you assign.

  1. In Designer, give each node that shows up in the interaction log a meaningful name (for example, Knowledge Search: Pricing)

  2. In Interaction Logs, switch the search mode to Search by Name

  3. Enter the node name to filter the table to every step that node has produced across all sessions and agents

Best Practices

  • Start from the failing step, not the response text: The trace tells you what the workflow actually did. The response is only the consequence.

  • Triage globally with Errors > Yes: When you do not yet know which session is affected, filter the view by errors and scan across agents and deployments

  • Paste an Event ID to pull a whole turn: When a bug report or log line gives you an event ID, the search returns every step that ran in that one workflow execution

  • Pivot back with View in Session: Once you have identified the problematic step, drop back into the conversation to see what the user saw and what came next

Last updated