# Integrated AI assistance

The integrated AI workflow extracts proposed graph records from text you supply. It uses your own OpenRouter key. It does not browse the web, fetch a pasted URL, query a registry, run tools or independently establish facts.

## Configure your test connection

**Prerequisites:** your own OpenRouter account/key and authorization to send the chosen synthetic text to the provider. A connection test and generation may consume provider usage.

1. Open **Settings** and find **OpenRouter test connection**.
2. Enter your key in **OpenRouter API key**, choose a **Model** from the app's current list, and choose **Save** or **Test connection**. Testing also saves the entered key and selected model locally.
3. Read the connection result. A saved key is not the same as a successfully tested connection.
4. Use **Clear key** to remove the saved key from this browser when finished.

The prototype stores that key in plain localStorage. Do not put credentials into Source text, Network notes, a shared snapshot, a prompt sent to another assistant or a recording. The application does not supply a shared Lage Bureau API key.

## Generate a reviewable proposal

In an open Network, choose **AI graph proposals**. Paste the exact material into **Source text**. This small exercise is authored fiction:

```text
Fictional documentation exercise.
LB Docs — Alex Example is a director of LB Docs — Sample Company GmbH.
The exercise does not state that Alex owns the company.
```

Inspect **Include selected Network context**. Only checked records are sent as context; selecting a relationship also includes its endpoint entities. The request contains supplied text and the selected record context, not a general browser research session. Choose **Generate proposal** only after deciding this is the material to send.

**Expected result:** a proposal to review, not new durable records. The exact suggestions depend on the selected model; the example is not a promise of a deterministic response.

## Review, apply and undo

1. In **Review proposal**, inspect proposed entities, relationships, Sources and Citations, their validation state and supporting excerpts.
2. Compare each relationship's excerpt with the submitted text. Check endpoints and direction. A syntactically valid proposal can still be an unhelpful interpretation.
3. Use **Accept proposal item** or **Reject proposal item**. A relationship needs valid accepted or existing endpoints; a citation needs its Source and target. Rejecting a dependency can block dependent items.
4. Leave **Analyst confidence after acceptance** at Unconfirmed unless you deliberately assess otherwise. **Model confidence** does not set it for you.
5. After reviewing the accepted set, choose **Apply accepted items**. The batch becomes one local, undoable edit. Check the save status and inspect the resulting records.
6. Use **Undo** if the accepted batch is wrong. Acceptance within the proposal dialog alone does not apply the batch.

**Accept all valid** is a convenience control, not an independent evidence review. Browser agents must follow the user's actual authorization and preserve that review decision. If generation, excerpt validation or dependency checks fail, read the error, correct the input or accepted subset, and try again deliberately. Do not bypass an invalid result or silently retry requests that may incur cost.