Models is where you assign an AI model to each function of your agent — task detection, task execution, embedding, rephrasing, ranking, response generation, and guardrail checks.
Overview
Its description reads: "Select AI models to configure agent functions. Use reasoning models for complex tasks or lightweight models for simple operations." It's a set of model pickers grouped by stage — Task Processing, Indexing, Pre-Processing, Response Generation, and Guardrails.
Note: There are no temperature, top-p, or token sliders here — those inference parameters belong to the model itself and are set in the Model Library when a model is registered. This screen is purely about which model does what. Changes save automatically — there's no Save button on this screen.
Which Functions Can I Assign a Model To?
Each function has its own Select Model picker:
Group | Function | Description |
|---|---|---|
Task Processing | Task Detection | Identifies which task the agent should run based on user input. |
Task Processing | Task Execution | Executes the selected task and generates step-by-step actions. |
Indexing | Embedding Model | Converts organized data into vector points to match against user queries. |
Pre-Processing | Query Rephrasing | Rephrases user queries for better retrieval accuracy. |
Pre-Processing | Ranking Model | Re-ranks retrieved chunks by relevance to improve result quality. |
Response Generation | Response Generation | Generates accurate and contextual responses based on retrieved knowledge. |
Guardrails | Hallucination Detection | Detects potential hallucinations or factual mismatches in generated responses. |
Guardrails | Content Filter | Filters inappropriate content in user inputs and model outputs. |
Guardrails | Denied Topic Detection | Identifies and blocks restricted topics in conversations. |
Each picker shows the model name and its provider icon. Use a stronger reasoning model where quality matters (task execution, response generation) and a lighter, faster model for simple checks (rephrasing, ranking, guardrail detection) to balance cost and speed.
Where Do the Available Models Come From?
The model list in every picker is your workspace's catalog of registered models, managed in the Model Library. It's not a fixed list — it reflects whatever models have been added and enabled there, filtered to the ones suited to each function (for example, only embedding models appear in the Embedding Model picker). Each function starts filled in with the workspace default model until you pick a different one. You can switch a model but not blank it out — every function must always have a model.
Except for Embedding Model and Ranking Model, each picker can also select a model router, which routes each call to the right model. Switch between the All, Routers, and Models tabs at the top of the list, or create a router from the list.
Where Do I Set Temperature, Max Tokens, or Reasoning Effort?
Not on this screen. Inference parameters — Temperature, Top P, Top K, Max Output Tokens, Frequency/Presence Penalty, and reasoning/thinking controls — are configured on the model itself in the Model Library when it's registered. The agent's Models section only selects which model to use.
Response-shaping options that are agent-level (structured output format, showing the agent's reasoning, language) live in the Response generation settings, not here.
Do Model Changes Save Automatically?
Yes. The Models section auto-saves the moment you change a picker — there's no Save button. If the screen is read-only (you're viewing a published version or lack edit access), the pickers are disabled.