> ## Documentation Index
> Fetch the complete documentation index at: https://docs.neuralfactory.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Interactions

> The sealed record of every agent turn and workflow run, with who asked, what ran, what it did and what it cost.

**Interactions** is where you look back at what an agent or a workflow did. An administrator uses it to answer questions such as "who asked the HR agent about salaries last week, what did it read, and what did it answer?". A Member uses it to review their own tasks and runs.

> [!NOTE]
>
> **Who:** Admins see the whole organization; Members see their own · \*\*Where:\*\***Governance** › **Interactions** in the sidebar · **Works on:** Enterprise and SaaS

The list holds every turn an agent answered and every workflow run that ended. Each row opens an interaction record: who asked and on whose behalf, which agent version or workflow revision ran, every step it took, who decided anything along the way, how it ended and what it cost.

A record is written once, when the turn or run ends, and sealed with a fingerprint. It cannot be edited afterwards, and it is checked again every time you open it. Deleting a chat does not delete its record.

## Who sees what

| Who | Sees |
| - | - |
| Admin | Every record of the organization, and the files in them |
| Member | Only the records of their own turns and of the workflow runs they started |
| Viewer (Enterprise) | Nothing: the page is not in their sidebar |
| A personal account (SaaS, no organization) | Their own records, including those made through their own API keys |

The same rule applies to the list, the record page and the download.

> [!NOTE]
>
> An Admin can preview and download the files in any record of their organization: the files a colleague attached to a chat, and the files a turn or run produced. Members and Viewers open only the files of their own chats. An Admin of another organization cannot open them. An API key opens only the files of its creator's own chats, even when it holds the Admin role.

## Find an interaction

Each row of the list is one agent turn or one workflow run.

* **Search request, subject, person…** finds rows whose request, answer, person's name or email, or channel contains the text.
* **All**, **Agent chats** and **Workflow runs** show both kinds of row, or only one.
* **Pick a date range** limits the list to the days you choose, counted in your time zone. **Clear dates** removes the range.
* Click a column header (Time, Who, Worker, Request, Cost) to sort by it.

Click a row to open its record. Hold Ctrl (Cmd on a Mac) or use the middle mouse button to open it in a new tab.

![The Interactions list with the search box, the All, Agent chats and Workflow runs tabs, and three completed agent turns showing who asked, the agent, the request, the duration, the cost and the Sealed record state.](https://mintcdn.com/neuralfactory/y-MI7aIFykaw0cY2/images/app/interactions.png?fit=max\&auto=format\&n=y-MI7aIFykaw0cY2\&q=85\&s=d22e15ff9bfba508c2cae035b334e039)

| Column | What it shows |
| - | - |
| **Time** | When it happened, in your browser's time zone. Hover the time to see it in UTC. |
| **Who** | The person who asked, or the API key that called. When a key or a schedule acted for someone, the row says **for** that person. The channel (for example Teams, API or Schedule) follows. |
| **Worker** | The agent and its version (`v3`), or the workflow and its revision (`rev 2`), with the number of subagents or delegations it used. |
| **Request** | What was asked, or the input of a workflow run. |
| **Outcome** | **Completed**, **Failed**, **Cancelled** or **Timed out**, with the number of tool errors and approvals, and the feedback the person gave. A dash means the outcome was not recorded (older rows). |
| **Duration** | The time it was working. Time spent waiting for a person (an approval or a question) is shown apart, as **+… waiting**. |
| **Cost** | What the turn cost. A dash means no cost. See [Cost on Enterprise and SaaS](#cost-on-enterprise-and-saas). |
| **Record** | **Sealed**, **Incomplete** or **No step record** (see [Record states](#record-states)). |

## Read a record

The top of the record shows the request (or the workflow's name and input), then **Worker**, **Started** (who started it, through which channel, and when), **Status**, **Duration** (with the waiting time beside it) and **Cost** (with the number of tokens).

![An interaction record of a question about annual leave, with Worker, Started, Status, Duration and Cost at the top, the Sealed and verified on open strip, the Trace tab with three steps on a time axis, and the Overview panel showing the request, the response and the cited source.](https://mintcdn.com/neuralfactory/y-MI7aIFykaw0cY2/images/app/interaction-record.png?fit=max\&auto=format\&n=y-MI7aIFykaw0cY2\&q=85\&s=fb9949b74eebe1bf905af3568b493548)

A finished workflow run also opens its record from the run's own page, with the **Record** button.

### Trace tab

**Trace** shows every step on a time axis: model calls, tool calls, knowledge-base retrieval, subagents, workflow steps, and decisions made by a person. Time spent waiting for a person is drawn apart from work.

* The chips above the steps (**All**, **Model**, **Tools**, **Retrieval**, **Subagents**, **Logic**, **Approvals**, **Skipped**, **Errors**) show one kind of step. A chip appears only when the record has such steps.
* **Find in steps** searches the steps.
* **Fold waiting time** shortens long waits on the time axis so the work stays readable.
* Times on the axis are offsets from the start. Hover a bar to see its start in UTC, its duration and its tokens.

Select a step to read it in the panel beside the trace. It shows what the model said, which tool it called with what input and what came back, which documents were retrieved, what a subagent was asked and returned, or who answered a question or decided an approval, when, and with what comment. A workflow step that did not run says why: another branch was taken, or the run ended or failed first.

When you select **Interaction** at the top of the trace, the panel shows the whole interaction. Its **Overview** lists the interaction's files under **Files produced**, where you can preview and download them.

A step's panel has **Overview**, **Input**, **Output** and **JSON** tabs. **Copy step link** copies a link that opens the record on that step. When a record fails, it opens on the first step that failed.

### Audit tab

**Audit** lists the facts a reviewer needs, as they were when the turn ended.

| Part | What it holds |
| - | - |
| Who, and on whose authority | The person or API key, their role and groups at the time, who a key or schedule acted for, and the channel. |
| What ran | The agent and its version or the workflow and its revision, a fingerprint of its configuration, the subagents, models, knowledge bases and tools used, and links to a run this one started or was started by. |
| Integrity | Whether the record was verified on open, when it was sealed, how many steps it holds, and its fingerprint. |
| Data | The subject the caller named, if any, and whether the content was stored. |

A later change of role, a renamed agent or an edited workflow never changes what an existing record says.

## Share or export

* **Copy link** in the record's header copies a link to this record. Anyone you send it to sees it only if they are allowed to (see [Who sees what](#who-sees-what)).
* **Export record** downloads the whole record as a JSON file: the record, every step with its full content, and a short summary of the integrity check made when the file was created.
* **Download CSV** on the list downloads every row you are allowed to see for the tab (**All**, **Agent chats** or **Workflow runs**) and the date range you picked, newest first. The search text does not narrow the download. Times in the file are in UTC.

> [accordion] What the CSV file contains
>
> One line per row, with the person's name and email, the request and the response, the time, the channel, the agent, the API key name, the knowledge bases, subagents and tools used, the feedback and its comment. Workflow runs add the workflow name, run id, run status, duration and how the run was started.

## Record states

The strip under the record's header says whether the record is intact.

| State | Meaning |
| - | - |
| **Sealed … · verified on open** | The record is complete and matches what was sealed. |
| **This record was changed after it was sealed.** | The check failed. The strip names the first step, or the part of the record, that no longer matches. |
| **Incomplete record.** | Sealing did not finish. The steps that were stored are still checked. |
| **No step record.** | The interaction is older than step-by-step records. Only the request, the response, the names of the tools used and the totals are kept. |

## Cost on Enterprise and SaaS

* **Enterprise**: the cost is the provider cost.
* **SaaS**: the cost is the price you are charged, including the platform margin.

## What is not in the list

> [accordion] Turns sent through the API without saving
>
> A turn sent through the API with `save_interactions` set to `false` is not listed. See the [API reference](https://docs.neuralfactory.ai/api-reference/introduction.md).

> [accordion] Runs that leave no record
>
> A run that fails before its first model call leaves no record. Neither does a workflow run the platform stops, for example for running past its maximum time.

> [accordion] How long records are kept
>
> Records are not deleted automatically.

## Related

* [Analytics](https://docs.neuralfactory.ai/user-guide/admin/analytics.md): totals and trends instead of single records.
* [Review agent activity](https://docs.neuralfactory.ai/user-guide/guides/review-agent-activity.md): a step-by-step guide.
* [Running a workflow](https://docs.neuralfactory.ai/user-guide/workflows/running-a-workflow.md)
* [Users and roles](https://docs.neuralfactory.ai/user-guide/concepts/users-and-roles.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.