LLM Web Search

Web search lets LLMs access real-time information from the internet when generating responses. Instead of relying solely on training data, the model can search the web to ground its answers with up-to-date facts, links, and sources.

How It Works

Enable web search by adding web_search_options to your chat completion request. When enabled, the model will:

Analyze the user’s query to determine if web search is needed.
Search the web for relevant, current information.
Generate a response grounded in the search results, often including source citations.

Basic Example

import requests

url = "https://api.edenai.run/v3/chat/completions"
headers = {
    "Authorization": "Bearer YOUR_API_KEY",
    "Content-Type": "application/json"
}

payload = {
    "model": "vertex/gemini-flash-latest",
    "messages": [
        {"role": "user", "content": "What are the latest developments in AI regulation in 2025?"}
    ],
    "web_search_options": {
        "search_context_size": "medium"
    }
}

response = requests.post(url, headers=headers, json=payload)
result = response.json()
print(result["choices"][0]["message"]["content"])

const url = 'https://api.edenai.run/v3/chat/completions';
const headers = {
  'Authorization': 'Bearer YOUR_API_KEY',
  'Content-Type': 'application/json'
};

const payload = {
  model: 'vertex/gemini-flash-latest',
  messages: [
    { role: 'user', content: 'What are the latest developments in AI regulation in 2025?' }
  ],
  web_search_options: {
    search_context_size: 'medium'
  }
};

const response = await fetch(url, {
  method: 'POST',
  headers: headers,
  body: JSON.stringify(payload)
});

const result = await response.json();
console.log(result.choices[0].message.content);

curl -X POST https://api.edenai.run/v3/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vertex/gemini-flash-latest",
    "messages": [
      {"role": "user", "content": "What are the latest developments in AI regulation in 2025?"}
    ],
    "web_search_options": {
      "search_context_size": "medium"
    }
  }'

Supported Models

Not all models support web search. Use the List LLM Models endpoint to find models where capabilities.supports_web_search is true:

import requests

response = requests.get("https://api.edenai.run/v3/models")

web_search_models = [
    model["id"]
    for model in response.json()["data"]
    if (model.get("capabilities") or {}).get("supports_web_search")
]
print(web_search_models)

On Qwen and DeepSeek, the models that search are qwen/qwen3.8-omni-flash and deepseek/deepseek-v4-pro. See Qwen Web Search and DeepSeek Web Search.

You can also browse models visually in the Eden AI model catalog and filter by capability, including web search.

OpenAI models: on /v3/chat/completions, web search only runs on the dedicated search variants (e.g. openai/gpt-5-search-api, openai/gpt-5-search-api-2025-10-14). Other OpenAI models either reject web_search_options with a 400 error or accept the request without performing any search. To use web search with regular OpenAI models, call them through the Responses API instead.

Parameters

Parameter	Type	Description
`web_search_options`	object	Options for web search integration. Omit to disable web search.
`web_search_options.search_context_size`	string	Amount of context retrieved from the web. Options: `"low"`, `"medium"`, `"high"`. Default: `"medium"`

Qwen Web Search

qwen/qwen3.8-omni-flash searches the web through Alibaba’s own search. On other Qwen models the request runs without search.

Turn it on with web_search_options on Chat Completions, or a web_search tool on Responses. An enable_search field in your request is ignored: Eden AI sets it.
Streaming only. Set stream to true. A search request without it is refused with 400.
Global only. Search runs on a Global-scope workspace, so it has no EU route. A call tagged @eu runs on the EU workspace and does not search. See EU data residency.
No search rate, no search. A model whose pricing carries no per-search rate answers 501.

Python

import requests

url = "https://api.edenai.run/v3/chat/completions"
headers = {
    "Authorization": "Bearer YOUR_API_KEY",
    "Content-Type": "application/json",
}

payload = {
    "model": "qwen/qwen3.8-omni-flash",
    "messages": [{"role": "user", "content": "What changed in EU AI rules this month?"}],
    "web_search_options": {"search_context_size": "medium"},
    "stream": True,
}

with requests.post(url, headers=headers, json=payload, stream=True) as response:
    for line in response.iter_lines():
        if line:
            print(line.decode())

Billing

A call is billed for its tokens plus the searches Alibaba reports, at the model’s per-search rate (web_search_cost_per_call).

Case	Billed
The model searched	The count Alibaba reports, times the per-search rate.
The model chose not to search, so Alibaba reports no search	No search fee.
The stream is cut before Alibaba reports its usage	A reservation of 10 searches.

DeepSeek Web Search

deepseek/deepseek-v4-pro searches the web through DeepSeek’s own search, on two endpoints.

Responses: add a web_search tool.
Messages (/v3/v1/messages): add Anthropic’s web_search server tool.

The model decides when to search. Web search on DeepSeek is not available on Chat Completions: a web_search_options field there does not make the model search. Request body for POST /v3/responses:

{
  "model": "deepseek/deepseek-v4-pro",
  "input": "What is the latest stable release of Rust?",
  "tools": [{"type": "web_search"}]
}

Each search appears in output as a web_search_call item, and action.queries lists the queries it ran. Request body for POST /v3/v1/messages:

{
  "model": "deepseek/deepseek-v4-pro",
  "max_tokens": 1024,
  "messages": [{"role": "user", "content": "What is the latest stable release of Rust?"}],
  "tools": [{"type": "web_search_20250305", "name": "web_search", "max_uses": 5}]
}

Each search appears as a server_tool_use block, and usage.server_tool_use.web_search_requests counts them.

Limits

Responses searches on deepseek/deepseek-v4-pro only. On deepseek/deepseek-flash DeepSeek ignores the tool and the request runs without search.
Chat Completions does not search. DeepSeek’s chat endpoint has no search parameter.
Data is processed in China. DeepSeek processes and stores requests in the People’s Republic of China, and DeepSeek search carries no eu tag, so it has no EU route. See EU data residency.
No search rate, no search. A model whose pricing carries no per-search rate answers 501 on Responses, and on Messages when the request carries a server search or fetch tool.

Billing

A call is billed for its tokens plus its searches, at the model’s per-search rate (web_search_cost_per_call in its pricing, listed by GET /v3/models), one query at a time. On Responses, each page the model opens counts as one query. The text it reads is counted in its input tokens.

Case	Billed
Responses, the model searched	The queries listed in each `web_search_call` item, plus one for each page it opened, times the per-search rate.
Messages, the model searched	`web_search_requests` (plus `web_fetch_requests`), times the per-search rate. Without that count, one per `server_tool_use` block.
The model chose not to search	No search fee.
A search count that cannot be read	At least 10 searches.
A Responses stream is cut before its final event	The larger of the queries it showed and 10.
A Messages stream is cut before its final usage	The larger of the search blocks it showed and the reservation below.

Every query the call ran is billed, including one that returns no results. Before the call, your balance is held for the searches as well as the tokens. Responses holds 10 searches. Messages holds the sum of the max_uses you set when every search tool states one, and 10 otherwise.

With System Messages

Combine web search with system messages to control how the model uses search results:

import requests

url = "https://api.edenai.run/v3/chat/completions"
headers = {
    "Authorization": "Bearer YOUR_API_KEY",
    "Content-Type": "application/json"
}

payload = {
    "model": "vertex/gemini-flash-latest",
    "messages": [
        {
            "role": "system",
            "content": "You are a research assistant. When answering, always cite your sources with URLs."
        },
        {
            "role": "user",
            "content": "What is the current stock price of Apple?"
        }
    ],
    "web_search_options": {
        "search_context_size": "high"
    }
}

response = requests.post(url, headers=headers, json=payload)
result = response.json()
print(result["choices"][0]["message"]["content"])

When to Use Web Search

Web search is useful when the user’s question involves:

Current events — news, stock prices, weather, sports scores
Recent releases — software versions, product launches, research papers
Factual lookups — statistics, regulations, schedules that change over time
Verification — checking claims against live sources

For questions about static knowledge (math, programming concepts, general reasoning), web search adds latency without much benefit. Only enable it when freshness matters.

Next Steps

Topic Extraction

Extract structured topics from text with Expert Models

V3 Documentation

Quick Start

Overview

LLMs

Expert Models

General

Organization

Data Governance

Integrations

How It Works

Basic Example

Supported Models

Parameters

Qwen Web Search

Billing

DeepSeek Web Search

Limits

Billing

With System Messages

When to Use Web Search

Next Steps

Topic Extraction

​How It Works

​Basic Example

​Supported Models

​Parameters

​Qwen Web Search

​Billing

​DeepSeek Web Search

​Limits

​Billing

​With System Messages

​When to Use Web Search

​Next Steps

Topic Extraction

How It Works

Basic Example

Supported Models

Parameters

Qwen Web Search

Billing

DeepSeek Web Search

Limits

Billing

With System Messages

When to Use Web Search

Next Steps