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

# Web tools

> Web tools available as workflow steps and through /v1/capabilities.

**1 tools.** Each one is deterministic: the same input gives the same output, and
nothing is inferred. Call one directly with
[`POST /v1/capabilities/{id}/run`](/api-reference/capabilities/run), or use it as a step in a
[workflow](/workflows/step-types).

Prices are on the [live catalogue](https://infery.ai/models) — they change, and a stale price here would be
worse than none.

| Tool                       | Does                                                                                                                | In → out   |
| -------------------------- | ------------------------------------------------------------------------------------------------------------------- | ---------- |
| [`web.search`](#websearch) | Search the web. Returns a grounded `answer` plus source links in `results`. Gateway picks provider by latency/cost. | any → text |

## `web.search`

**any → text** · Billed per request

Search the web and return ranked results. The gateway routes to the best available provider (Brave, Gemini, OpenAI, or Anthropic) based on latency and cost.

### When to use

* **Research queries** — retrieve up-to-date information from the live web.
* **Fact checking** — ground LLM responses in current sources.
* **News & events** — look up recent happenings beyond a model's training cutoff.

### Inputs & outputs

|                      |                                                 |
| -------------------- | ----------------------------------------------- |
| **Input modality**   | None (query is a parameter)                     |
| **Output modality**  | Text (a grounded answer plus a list of sources) |
| **Max query length** | 500 characters                                  |
| **Max results**      | 50                                              |

### Output: which field carries the content

The step returns `{ answer, results, citations, metadata }`.

**Bind `answer`, not `results`.** `results` is a list of *sources* — url, title,
rank, domain. It is not the text of those pages. `answer` is the provider's
grounded, synthesised answer to the query, and it is what a downstream model
step can summarise.

| Field                      | What it holds                                                                                               |
| -------------------------- | ----------------------------------------------------------------------------------------------------------- |
| `answer`                   | The grounded answer. `null` only on Brave, which does not synthesise.                                       |
| `results[]`                | `{ url, title, rank, domain, snippet? }` — the sources.                                                     |
| `results[].snippet`        | An extract of the source page, **when the provider returns one**. Absent otherwise — never an empty string. |
| `citations[]`              | `{ url, title, quote?, start_index?, end_index? }` — which source backs which part of `answer`.             |
| `metadata.snippet_support` | Whether `snippet` can be populated at all on the provider that served this request.                         |

`snippet` depends on the provider the gateway routed to, which you cannot choose
and cannot see in advance, so `metadata.snippet_support` states it in the output:

| `snippet_support` | Meaning                                                       |
| ----------------- | ------------------------------------------------------------- |
| `full`            | Every result carries a page extract.                          |
| `cited_only`      | Only the results the answer cited carry one.                  |
| `none`            | The provider's API returns no page text at all. Use `answer`. |

A binding that must work whatever the router picks should use the fallback
operator: `${steps.search.output.answer | steps.search.output.results}` — it
takes the grounded answer where there is one and the (snippet-bearing) result
list where there is not. This is what the `web-research-summary` template does.

### Parameters

| Param         | Required | Description                                       |
| ------------- | -------- | ------------------------------------------------- |
| `query`       | yes      | Search query string (1–500 chars)                 |
| `max_results` | no       | Number of results to return (1–50). Default: `10` |

These are the only parameters accepted; any other key is rejected.

### Examples

#### Basic web search

```json theme={null}
{
  "type": "web.search",
  "input": {},
  "params": { "query": "best LLMs 2026", "max_results": 10 }
}
```

#### Non-English search

Results follow the language of the query — there is no locale parameter.

```json theme={null}
{
  "type": "web.search",
  "input": {},
  "params": { "query": "Bundesliga Ergebnisse", "max_results": 5 }
}
```

### Pricing

Billed per search request. The gateway routes to the most appropriate provider; see pricing dashboard for current rates.

### Related capabilities

* `code.run_python` — process or summarise search results with inline Python
* `code.run_node` — transform result data with Node.js

### Examples

**Basic search**

```json theme={null}
{
  "type": "web.search",
  "params": {
    "query": "best LLMs 2026",
    "max_results": 10
  }
}
```