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

# Model selection

> Choose Sofie chat models, set a default, and understand how model choice affects speed, reasoning, and context.

The model selector lets you choose which AI model Sofie uses for a chat. Model choice can affect speed, reasoning depth, context capacity, response style, and how well Sofie handles long or source-heavy work.

Your organization controls which models appear. You may not see every model described here.

<Warning>
  Model choice does not replace source review. For regulated, quality, clinical, or controlled work, verify sources, calculations, conclusions, and final language yourself.
</Warning>

## Open the model selector

<Steps>
  <Step title="Open chat">
    Start a new chat or open an existing conversation.
  </Step>

  <Step title="Open the model menu">
    Click the model name in the chat input.
  </Step>

  <Step title="Search or browse">
    Search models by name, or browse the **Recommended** section and provider groups.
  </Step>

  <Step title="Select a model">
    Click a model to use it for the chat.
  </Step>
</Steps>

The selector can show:

* A **Recommended** section.
* Provider groups such as Anthropic, Google, OpenAI, or xAI.
* Model names.
* Short model descriptions.
* A checkmark for the current model.
* A star control for your default model.

## Set your default model

Use the star in the model selector to set your default model.

<Steps>
  <Step title="Open the model selector">
    Click the current model name in chat.
  </Step>

  <Step title="Find the model">
    Search or browse to the model you want to use by default.
  </Step>

  <Step title="Click the star">
    Click the star next to the model.
  </Step>

  <Step title="Confirm future chats">
    New chats use your default model when available.
  </Step>
</Steps>

If a model is no longer available, Sofie may fall back to the current system default.

## Choose by task

Use the task to guide model choice.

| Task                                                                   | Model direction                                                                 |
| ---------------------------------------------------------------------- | ------------------------------------------------------------------------------- |
| Quick rewrite, short summary, title, or classification                 | Use your default recommended model or a faster model.                           |
| Long document review                                                   | Use a model suited to longer context and keep sources focused.                  |
| Deviation investigation, CAPA, validation, URS, or regulatory drafting | Use a stronger reasoning model and provide source context.                      |
| Brainstorming or first-pass outline                                    | Use a faster or balanced model, then switch if deeper review is needed.         |
| Source comparison with many documents                                  | Use a model with strong reasoning and manage the context window carefully.      |
| Orchestration design                                                   | Use a stronger reasoning model when task logic, handoffs, or tests are complex. |

## Understand common tradeoffs

Model choice usually involves tradeoffs.

| Tradeoff             | What it means                                                                                 |
| -------------------- | --------------------------------------------------------------------------------------------- |
| Speed                | Faster models can be better for short tasks and iteration.                                    |
| Reasoning            | Stronger reasoning models can be better for complex analysis, tradeoffs, and multi-step work. |
| Context              | Larger context windows can handle more material, but focused source selection still matters.  |
| Cost or availability | Your organization may limit which models are available.                                       |
| Style                | Different models may write, summarize, or structure answers differently.                      |

<Tip>
  When a task fails because the source set is too broad, changing models may help, but narrowing context usually helps more.
</Tip>

## Recommended models

The **Recommended** section is the fastest place to start. It can include models chosen by your organization or Sofie defaults for general chat use.

Use a recommended model when:

* You are starting a new chat.
* You are unsure which model to pick.
* You need a balanced default for everyday Sofie work.
* The task is not unusually long, complex, or source-heavy.

Switch away from the recommended default when you have a clear reason, such as deeper reasoning, longer context, or faster iteration.

## Model choice and context windows

Each model has a context window: the amount of information it can consider in one response. The context indicator in chat shows approximate usage for the current model.

Model choice can affect:

* How much chat history and selected context fits.
* Whether long source material needs more summarization.
* How much room remains for attachments, search results, and generated output.
* When a long chat should be summarized or restarted.

Even with a large context window, do not overload the chat with unrelated sources. Focused context usually produces better answers than a large, mixed source set.

For detailed context guidance, see [Context windows and summarization](/chat/context-windows-and-summarization).

## Switch models during work

You can change models when the work changes.

Common patterns:

| Situation                                       | Move                                                               |
| ----------------------------------------------- | ------------------------------------------------------------------ |
| You are drafting rough options                  | Use a faster or balanced model.                                    |
| You are checking a complex investigation        | Switch to stronger reasoning.                                      |
| You are comparing long source material          | Switch to a model suited to longer context and narrow the sources. |
| You are editing tone after analysis is complete | Switch back to a faster or default model.                          |

When you switch models, restate any important constraints if the task is sensitive:

```text theme={null}
Continue with the same source rules: use only the selected Workspace and attached protocol. Separate facts, assumptions, source gaps, and SME questions.
```

## Model choice and Plan Mode

Use [Plan Mode](/chat/plan-mode) when the model needs to coordinate several steps, tools, sources, or review points.

Model choice matters in Plan Mode because Sofie may need to:

* Understand the objective.
* Choose the right source path.
* Sequence tasks.
* Decide where review is needed.
* Keep constraints consistent across steps.

For complex life sciences work, pair a stronger reasoning model with Plan Mode and clear source rules.

Example:

```text theme={null}
Use Plan Mode with a stronger reasoning model. Search the Workspace for URS and validation protocol sources, identify gaps, then create a CoDraft outline only after I approve the gap table.
```

## Model choice and Orchestrations

Chat model selection affects the current chat. Orchestration agents may also have model settings in the Orchestration editor when your organization offers choices.

Use stronger reasoning for Orchestration agents that:

* Interpret source-heavy evidence.
* Decide task routing.
* Compare conflicting information.
* Produce structured outputs used by later agents.
* Draft quality, validation, or regulatory text.

Use faster models for Orchestration agents that:

* Rename items.
* Classify simple inputs.
* Reformat known fields.
* Produce short summaries from already-clean source context.

For deeper workflow guidance, see [Use Orchestrations intelligently](/orchestrate/use-orchestrations-intelligently).

## Troubleshoot model-related issues

| Issue                        | Try                                                                            |
| ---------------------------- | ------------------------------------------------------------------------------ |
| Answer is too shallow        | Add sources, use a stronger reasoning model, and ask for assumptions and gaps. |
| Answer misses source details | Narrow context, use Workspace search, and ask for citations.                   |
| Chat feels slow              | Use a faster model for simple edits or short summaries.                        |
| Context indicator is high    | Narrow sources, summarize, or start a new chat.                                |
| Output format drifts         | Use a Saved prompt, Directive, or structured table request.                    |
| Model is missing             | Ask your administrator whether that model is enabled for your organization.    |

## Good prompts for model-sensitive work

<Accordion title="Use stronger reasoning for investigation work">
  ```text theme={null}
  Use a stronger reasoning model for this deviation review. Use only the selected Workspace and attached summary. Return facts, assumptions, missing evidence, source gaps, and SME questions. Do not conclude root cause.
  ```
</Accordion>

<Accordion title="Keep context focused">
  ```text theme={null}
  Before answering, list the selected sources you will use. If the source set is too broad for the current model context, recommend what to remove or summarize.
  ```
</Accordion>

<Accordion title="Switch to faster drafting">
  ```text theme={null}
  Now that the analysis is approved, use the current conclusions to draft a concise CoDraft section. Do not add new claims.
  ```
</Accordion>

## Related guides

* [Chat with Sofie](/chat/chat-with-sofie)
* [Context windows and summarization](/chat/context-windows-and-summarization)
* [Use Plan Mode](/chat/plan-mode)
* [Saved prompts](/prompts/saved-prompts)