Model
Choose the LLM that processes an IDP template: Roberty's built-in LLM or an external connection to OpenAI, Anthropic or Google Gemini.
Model
The Model section of the settings sets which LLM processes the template. Every LLM stage in the pipeline uses this same model:
- identification;
- LLM OCR, in extraction;
- the LLM steps of post-processing.
There is no way to pick a different model for a specific stage. Changes are saved automatically to the draft and reach production when you publish a version. See Versions.
LLM stages need the generative AI feature enabled on the workspace. If they show as locked, contact your account manager to have the feature turned on.
Provider
| Option | Description |
|---|---|
| Roberty's built-in LLM | The default. Ready to use, no setup. |
| External connection | Uses your own OpenAI, Anthropic or Google account, through a workspace connection. |
External connection
With External connection, choose:
| Field | Description |
|---|---|
| Service | OpenAI, Anthropic or Google Gemini. Changing the service clears the chosen connection and model. |
| Connection | A workspace connection to the chosen service. Changing the connection clears the chosen model. The Manage workspace connections link opens the workspace connections in another tab. |
| Model | One of the models available to the connection, listed from the account itself, with search. When the provider reports it, the context window shows next to the name. |
Connections are created in the workspace settings, by administrators only. See Connections, and the OpenAI, Anthropic and Gemini pages.
Model parameters
With an external connection, you also tune how the model generates its answers:
| Parameter | Range | Description |
|---|---|---|
| Temperature | 0 to 2 | Higher is more creative; lower is more predictable. |
| Top P | 0 to 1 | An alternative to temperature for controlling how much answers vary. |
| Max tokens | 1 to 32,000 | The maximum size of each model answer. |
When both temperature and Top P are set, temperature wins and Top P is ignored. Identification always uses temperature 0, whatever value is set here.
Imported templates
When a template is imported, the default model's connection is checked like the others: if it doesn't exist in the workspace, choose another one during the import or later, in this section. See General.