Research on the web

Choose search capabilities and separate search results from verified source content.

Mellow separates discovering a page from reading its contents. Search returns candidates; extraction reads the selected sources. This helps an agent support an answer with material it actually retrieved instead of treating a search snippet as a complete article.

Start with a source-oriented task

Enable Web Search for the agent, then ask a question that identifies what evidence is needed:

Find the official release notes for the version I named. Read the compatibility section and summarize the changes that affect this Mac.

A useful result names the sources and distinguishes retrieved facts from inference. Search availability depends on your configured providers and network. A local model can still make network requests through search tools.

Choose the search route

Open Web Search settings to review enabled providers. Mellow supports configured API providers and keyless fallback paths. A hosted search route is usable only when the relevant hosted service is configured and available; Cloud workspace sign-in alone does not guarantee that route.

Add a provider using its secure API-key field, test with a narrow query, and inspect which provider returned results. Provider pricing and free allowances change, so consult that provider's current plan rather than relying on a fixed allowance in this guide.

A fallback can return different coverage or ranking. If a specialist provider fails and a general fallback succeeds, evaluate whether the result still meets your task's requirements.

Search and extraction tools

Extraction may return readable article text or preserve structured CSV, TSV, or JSON data. Large structured results can be passed through a data reference for follow-up tools instead of filling the model's context with every row.

Refine a query

web_search requires query. The source schema also supports max_results, time_range, site, filetype, region, offset, and provider-dependent category options. Use these to narrow evidence rather than adding vague adjectives to the question.

• Restrict site when an official source matters.
• Use a date range when old documentation could mislead.
• Request a file type when looking for a report or dataset.
• Broaden one constraint at a time if the result is empty.

Availability of news or image categories depends on the enabled providers. A result label should be checked before assuming a provider honored every requested filter.

Verify before relying on the answer

Open or extract important sources and check their date, version, and relevance. A page can discuss a different product version, quote an outdated policy, or be unavailable behind a sign-in. Ask the agent to say when extraction failed instead of presenting a snippet as full-page evidence.

For a long investigation, keep the original question and source criteria in the chat. Repeatedly rephrasing a query without reading any result usually adds little confidence.

Troubleshooting

No results: remove the most restrictive domain, time, or file filter; inspect attempted providers; test the provider directly through its settings.

Search works but reading fails: the destination may block automated retrieval, require authentication, or render content interactively. Use Browser Use if suitable.

A key is rejected: check the selected provider, credential, and account quota. Do not paste the key into the question.

Only fallback results appear: inspect the preferred provider's error before assuming it supplied the answer.

A data source is too large: ask for the relevant fields or use the structured-data follow-up workflow rather than copying the entire payload into chat.

Search sends your query to the serving provider, and extraction contacts destination sites. Use Privacy Filter for supported model-bound text and Security to understand the wider data paths.