Native web search is a toggle on the AI Model step that lets the model query the live web as part of generating its response. It is a capability built into the model provider’s API — not a connector, not an MCP tool, and not a separate credential. When the toggle is on, the model decides when to search, what to search for, and how to incorporate the results. This is different from tool-based search integrations like Brave Search or Firecrawl, where you wire a tool step into your agent workflow and control when it runs. Native web search is fully model-driven.Documentation Index
Fetch the complete documentation index at: https://explore.airia.com/llms.txt
Use this file to discover all available pages before exploring further.
Quick reference
| Setting | Default | What it does |
|---|---|---|
| Web Search | Off | Allows the model to query the live web before generating its response |
What is native web search?
When Web Search is enabled, the model’s host (Anthropic, OpenAI, or Google) performs a search on the model’s behalf during inference. The search results are injected into the model’s context before it generates a response, grounding the answer in current information. From your agent’s perspective, you enable a toggle. The model handles everything else — query construction, result retrieval, and synthesis. You do not need to write a prompt that asks the model to search; it will search when it judges that the question requires current information. Depending on the provider, the model’s response may include citations or source references inline.Supported models
The Web Search toggle appears only on models that support this capability. The full list of supported models is below.- Anthropic (17 models)
- OpenAI (5 models)
- Google (4 models)
| Model |
|---|
| Claude 3.5 Haiku Latest |
| Claude 3.5 Haiku Snapshot October 2024 |
| Claude 3.5 Sonnet Latest |
| Claude 3.5 Sonnet Snapshot October 2024 |
| Claude 3.7 Sonnet Latest |
| Claude 3.7 Sonnet Snapshot February 2025 |
| Claude Sonnet 4 |
| Claude Sonnet 4 Snapshot May 2025 |
| Claude Opus 4 |
| Claude Opus 4 Snapshot May 2025 |
| Claude Opus 4.1 |
| Claude Sonnet 4.5 |
| Claude Haiku 4.5 |
| Claude Opus 4.5 |
| Claude Sonnet 4.6 |
| Claude Opus 4.6 |
| Claude Opus 4.7 |
If you do not see the Web Search toggle in the Model panel, your currently selected model does not support native web search. Switch to one of the models listed above.
Enabling web search
- Open your agent in Agent Studio
- Select the AI Model step
- Open the Model tab in the right-hand panel
- Scroll below Temperature to find the Web Search toggle
- Toggle it on
When to enable / disable
- When to enable
- When to disable
- Research and Q&A agents that need answers about current events, prices, product versions, or recent releases
- Agents that must cite sources or provide verifiable, up-to-date facts
- Customer-facing assistants where stale training data would produce incorrect answers
- Monitoring or briefing agents that summarise what is happening right now
What the model searches — and what it receives
The model determines on its own when a search is needed and what query to issue. You do not write the search query and it is not visible to you in execution logs. The provider’s infrastructure runs the search, retrieves results, and injects them into the model’s context before it generates a response. What this means in practice:- You do not control the query. The model infers it from your prompt and the user’s message.
- Results are not exposed as a step output. You cannot pipe them into other steps or inspect them directly. Use the Debug tab to review the model’s full context if you need to understand what it retrieved.
- Citations may appear inline. Depending on the provider, the model may include source URLs or reference markers in its response. This is provider-specific behaviour and cannot be configured from Airia.
Considerations
Latency
Each web search adds a round-trip to the provider’s search infrastructure before the model generates text. Expect approximately 1–3 seconds of additional latency per execution. For real-time conversational agents this is usually acceptable; for high-volume batch pipelines it may not be.Token usage
Search results are injected into the model’s context window and count toward your token consumption. Responses from a web-search-enabled model typically cost more tokens than the same prompt without search. Monitor token usage in the Feeds view if cost is a concern.Privacy
Determinism
Web search results vary by time, query phrasing, and provider index freshness. Two identical prompts sent minutes apart may produce different answers. If reproducible outputs are required — for automated testing, audit trails, or compliance workflows — keep Web Search off.Native web search vs. search tool integrations
| Native Web Search | Brave Search / Firecrawl (MCP tools) | |
|---|---|---|
| Setup | Toggle only — no credentials needed | MCP tool added to agent + API key required |
| Control | Model decides when and what to search | You control when the tool runs in your workflow |
| Query visibility | Not inspectable | Tool input and output appear in execution logs |
| Results in outputs | Synthesised into the model response | Raw results piped into model context as a step output |
| Citation format | Provider-formatted, embedded inline | Depends on how you prompt the model to use the results |
| Best for | General research, real-time Q&A, conversational agents | Structured crawling, specific URLs, scraping, high-volume pipelines |
Troubleshooting
The Web Search toggle is not visible
The Web Search toggle is not visible
The toggle only appears for models that support native web search. If you don’t see it, your currently selected model does not support this feature. Open the model selector and switch to one of the supported models listed in the Supported models section above.
Responses are slower than expected
Responses are slower than expected
Native web search adds latency for each search query the model issues. If speed is critical, disable the toggle and consider using a Data Search step against a pre-indexed knowledge source, or a Brave Search MCP tool where you control when the search fires.
Model is not using current information despite the toggle being on
Model is not using current information despite the toggle being on
The model searches selectively — it only issues a query when it judges that current information is needed. If your prompt is phrased in a way that sounds like a general knowledge question, the model may answer from training data. Rephrase to make the recency requirement explicit (e.g., “What is the current price of X as of today?”). Also verify the toggle is saved by reopening the Model panel.
Unexpected token costs after enabling web search
Unexpected token costs after enabling web search
Search results injected by the model count as input tokens. Disable Web Search on any AI Model step that does not need real-time data. Use the Debug tab on a recent execution to inspect the full context sent to the model and identify how many tokens the search results consumed.
Sensitive data may be leaving my environment
Sensitive data may be leaving my environment
When Web Search is on, the model may issue queries constructed from your prompt content. If your agent handles confidential or personally identifiable information, disable Web Search and use an air-gapped retrieval approach (e.g., a Data Search step against your own knowledge base) instead.