Configuring model and data sources

To ensure that your agent delivers accurate results, it is crucial to select the right model and connect relevant data sources. These data sources form the agent's knowledge base.

The platform lets you configure different models and data source types for each agent. You can upload files, connect to cloud storage, or link to website URLs that can be indexed. The agent uses this information to quickly find and return relevant answers to your queries. Your agent can scan through all this information in no time to find solutions to your queries.

Organization model settings

Organization admins (and anyone with Organization Settings edit access) control which language models your organization can use and—when model restriction is on—which surfaces each model appears on.

Open organization model settings

1. In the left sidebar, open Organization > Models (or go to https://devs.ai/org-models).
2. Review the Large Language Models and Image Models tables.

Restrict which models are available

1. Turn on Restrict available Models if you want an explicit allowlist instead of every new model being enabled automatically for your organization.
2. Select which models are enabled org-wide and set default chat and image models.
3. Click Save.

When restriction is off, all models your plan supports are generally available everywhere. When restriction is on, only checked models are eligible—and you can narrow them further per surface (see below).

Per-surface model availability

After you turn on Restrict available Models, you can control where each allowed model may be used. Open a model’s configuration drawer (from the Models page) to enable or disable it independently for each surface:

Surface	Where it applies
Conversational Chat	Default homepage chat and general conversations that are not tied to a custom agent
Custom Agents	Model pickers when you create or edit an agent, and agent-backed chats
App Builder	Model selection while building or iterating on apps
Routines	Model selection for scheduled agent runs

For example, you might allow a fast, low-cost model for everyday chat but reserve a premium reasoning model for Custom Agents only—or keep experimental models out of App Builder until you are ready.

Surface-aware pickers list only models that are both on your organization allowlist and enabled for the surface you are working in. If a model appears in chat but not when you edit an agent, an admin has likely disabled it for Custom Agents (or the reverse).

If the org default model is disabled for a given surface, the product falls back to another model that is enabled on that surface.

📝 Note: Organizations that enabled model restriction before per-surface controls shipped treat every allowlisted model as enabled on all surfaces until an admin configures surface toggles.

Select a language model (LLM)

When you create a new agent or edit an existing one, click Model to open the catalog. What you can choose depends on your organization’s plan, enabled providers, region, and any organization model settings your admin has configured. The picker lists only models enabled for agents that you are allowed to use. Models differ in speed, cost, context size, and quality—use the labels in the product as the source of truth for what is available to you right now.

Azure OpenAI (enterprise GPT routes)

Many GPT-family models are offered through Azure OpenAI for organizations that use that deployment. Those routes are the ones that carry enterprise-grade privacy commitments where your agreement with Microsoft applies. For models that expose reasoning controls, Devs.ai sends only combinations the Azure endpoint accepts, so chats are less likely to fail because of invalid reasoning settings.

Organizations that route xAI traffic through Azure AI Foundry may also see Grok 4.3 (Azure) in the catalog, listed alongside other Grok models. That route is deployment-specific—if you expect Azure-hosted Grok but do not see it, ask your organization admin or account team whether Foundry is configured for your tenant.

Grok (xAI)

Grok 4.3 is available on direct xAI routes wherever Grok models appear in your picker. Catalog cards summarize tool calling, web search, vision, and context limits for the variants your organization enables.

Google (Gemini)

Google Gemini models in the catalog include the Gemini 3.x families, with native web search on supported variants. Gemini 3.5 Flash supports tool calling, so agents and chats that rely on tools (MCP connectors, built-in actions, and similar) can use this fast model where your organization has it enabled.

Auto (conversational chat)

Auto is available in conversational chat. For each turn, the platform picks a concrete model suited to your message while keeping Auto selected on the chat. Chats on Auto can attach the image generation tool when your request calls for a generated image, subject to organization tool policies and image model settings.

Kimi (Moonshot)

When Kimi 2.6 and related Kimi models appear for your organization, the catalog lists a 256K context window so expectations line up with provider limits. If web search is turned on for your environment and the model supports it, the agent can answer using web search in the chat alongside its other tools—the same org and agent settings that govern search for other models still apply.

Other providers

Depending on what your administrators enable, the picker may include models from OpenAI (direct), Anthropic, Google, Perplexity, xAI, Cohere, Meta, Mistral, and others. The catalog also adds newer families over time—for example GPT-5.6 (Sol, Terra, and Luna) and Grok 4.5 for chat and App Builder. Each card summarizes capabilities such as vision, tools, reasoning, and approximate context. Use the picker as the source of truth for what your organization can access today.

After you select a model, you can adjust parameters such as temperature and token limits where the model supports them.

Keep in mind that LLMs still struggle with precise math unless you give them structured tools or verification steps. If an agent fails or times out repeatedly, try a different model or simplify the task; for Azure-hosted GPT models, rare failures tied to reasoning configuration should be less common than in earlier releases.

Data sources

After selecting a model, connect different data sources to build a knowledge base for your agent to refer to.

• Previous Data - reuse data sources added for other agents.
• Upload Files - upload files directly from your computer, such as documents, spreadsheets, or PDFs. These files are indexed and used by the agent when responding to queries.
• Website URLs - enter URLs of websites or domains you want to index. The agent can scan and reference content from these pages.
• Google Drive — connect your Google account to import files and folders from Google Drive, including shared content (see Google Drive folders below).
• Microsoft OneDrive - connect your Microsoft account to import data from your Microsoft OneDrive.
• Confluence and JIRA - connect your Atlassian account to configure Confluence and JIRA to allow the agent to use issues, tickets, and documentation.
• Github - connect GitHub repositories so the agent can reference code, issues, and related documentation.

When using cloud storage or an online data source, you can set a frequency for data refresh. This controls how often data is indexed or pages are refreshed.

Large or complex data sources may affect response time. It is important to test your agent after making changes. Clear and focused data sources produce better results than large, unstructured collections.

Google Drive folders

You can link whole folders—not only individual files—from Google Drive to an agent’s knowledge base. This includes folders you own, folders in Shared with me, and folders in Shared drives.

Organization administrators must complete Google Drive integration before builders can connect personal Google accounts in the agent editor.

Browse and select folders

1. Open the agent editor and go to the Knowledge tab (or the data source step).
2. Choose Google Drive and connect or select your Google account if prompted.
3. Open the folder browser.
4. Pick a source list:
   • My Drive + Shared with me — Folders and files from your drive and items shared with you.
   • Shared drives — Team drives your account can access.
5. Expand folders to navigate, then select one or more folders to link.
6. Submit the selection. The agent creates data sources and indexes content in the background.

You can select multiple folders in one pass when you need several shared locations.

Paste a folder link

Alternatively, paste a Google Drive folder URL (for example a link that contains /drive/folders/ or /folders/). When you are signed in with Google Drive, permission-gated shared folders index using your account instead of requiring a public link.

File links pasted in the same field continue to work as before and index as single files.

What appears in the knowledge list

After a folder is linked:

• Files inside the folder are indexed for the agent to retrieve.
• Immediate subfolders appear as organizational rows in the knowledge list so you can see how content is grouped.
• Deeper subfolders and files are added as the background scan progresses.

Folder rows in the knowledge list are for navigation only. To link a different folder, use the folder browser or paste-link flow again—not the knowledge list itself.

Tips

• Reconnect your Google account if shared drives or team folders are missing after a permission change.
• Very large folders take longer to index. Test the agent after the initial scan completes.
• For SharePoint team sites, use the SharePoint data source flow instead.

After configuring datasources, you can set up additional actions and link tools to your agent.