Search Engine Providers

Each Search Engine wraps a single provider. The provider is selected with the provider field, and configured through config.connection (endpoint and credentials) and config.options (provider-specific options).

RAG knowledge base

The config.connection block below applies to the web providers. The rag provider is configured differently — it has no connection, and instead references an embedding store and an embedding model.

The config.connection block is common to all web providers:

Field	Type	Default	Description
config.connection.base_url	string	provider default	The provider base URL (or your instance URL for SearXNG)
config.connection.token	string	—	API key/token. Supports vault references and comma-separated rotation. Ignored by providers that don't require auth (SearXNG, DuckDuckGo).
config.connection.timeout	number	30000	Request timeout in milliseconds

The normalized response is always { provider, query, answer?, results: [{ title, url, snippet, score?, published_date? }] }.

Staan.ai (Qwant) 🇫🇷 🇪🇺

Sovereign European web search. POST https://api.staan.ai/v2/search/web, authenticated with a Bearer token.

Field	Type	Default	Description
config.connection.base_url	string	https://api.staan.ai	Staan base URL
config.options.market	string	fr-FR	Market / locale (e.g. fr-FR, en-us, de-de)
config.options.count	number	10	Number of results per page
config.options.min_score	number	—	Minimum reranker score (0–1)
config.options.extra_snippets	boolean	—	Return additional scored chunks per result
config.options.full_content	string	—	markdown or html to fetch full page content

{
  "id": "search-engine_xxxxxxxxx",
  "name": "Staan.ai",
  "provider": "staan",
  "config": {
    "connection": { "base_url": "https://api.staan.ai", "token": "${vault://local/STAAN_API_KEY}", "timeout": 30000 },
    "options": { "market": "fr-FR", "count": 10 }
  },
  "kind": "ai-gateway.extensions.cloud-apim.com/SearchEngine"
}

Tavily

Search API designed for LLM/RAG. POST https://api.tavily.com/search, Bearer token. Can return a synthetic answer.

Field	Type	Default	Description
config.connection.base_url	string	https://api.tavily.com	Tavily base URL
config.options.max_results	number	5	Number of results (0–20)
config.options.search_depth	string	basic	basic or advanced
config.options.topic	string	—	general, news, or finance
config.options.include_answer	boolean	—	Ask Tavily for an LLM-generated answer (returned in answer)

{
  "id": "search-engine_xxxxxxxxx",
  "name": "Tavily",
  "provider": "tavily",
  "config": {
    "connection": { "base_url": "https://api.tavily.com", "token": "${vault://local/TAVILY_API_KEY}", "timeout": 30000 },
    "options": { "max_results": 5, "search_depth": "basic", "include_answer": false }
  },
  "kind": "ai-gateway.extensions.cloud-apim.com/SearchEngine"
}

Brave Search

GET https://api.search.brave.com/res/v1/web/search, authenticated with the X-Subscription-Token header (set through connection.token).

Field	Type	Default	Description
config.connection.base_url	string	https://api.search.brave.com	Brave base URL
config.options.country	string	—	2-letter country code
config.options.search_lang	string	—	Content language (ISO 639-1)
config.options.safesearch	string	—	off, moderate, or strict

{
  "id": "search-engine_xxxxxxxxx",
  "name": "Brave Search",
  "provider": "brave",
  "config": {
    "connection": { "base_url": "https://api.search.brave.com", "token": "${vault://local/BRAVE_API_KEY}", "timeout": 30000 },
    "options": {}
  },
  "kind": "ai-gateway.extensions.cloud-apim.com/SearchEngine"
}

SearXNG

Open-source metasearch engine — great for on-prem and sovereignty. GET <instance>/search?format=json. No API key: point base_url at your SearXNG instance.

Field	Type	Default	Description
config.connection.base_url	string	http://localhost:8080	Your SearXNG instance URL
config.options.engines	string	—	Comma-separated list of upstream engines (e.g. google,bing,duckduckgo)
config.options.categories	string	—	Search categories (e.g. general)
config.options.language	string	—	Language code

{
  "id": "search-engine_xxxxxxxxx",
  "name": "SearXNG",
  "provider": "searxng",
  "config": {
    "connection": { "base_url": "https://searxng.internal.example.com", "token": "", "timeout": 30000 },
    "options": { "language": "fr" }
  },
  "kind": "ai-gateway.extensions.cloud-apim.com/SearchEngine"
}

Google Custom Search

Google Programmable Search Engine. GET https://www.googleapis.com/customsearch/v1. Authentication uses an API key (set through connection.token) and a search engine id cx.

Field	Type	Default	Description
config.connection.base_url	string	https://www.googleapis.com	Google APIs base URL
config.connection.token	string	—	The Google API key (sent as the key query param)
config.options.cx	string	—	The Programmable Search Engine id (required)
config.options.lr	string	—	Language restrict (e.g. lang_fr)

{
  "id": "search-engine_xxxxxxxxx",
  "name": "Google Custom Search",
  "provider": "google",
  "config": {
    "connection": { "base_url": "https://www.googleapis.com", "token": "${vault://local/GOOGLE_CSE_KEY}", "timeout": 30000 },
    "options": { "cx": "0123456789abcdef0" }
  },
  "kind": "ai-gateway.extensions.cloud-apim.com/SearchEngine"
}

SearchApi

SERP aggregator. GET https://www.searchapi.io/api/v1/search, Bearer token. Defaults to the google engine.

Field	Type	Default	Description
config.connection.base_url	string	https://www.searchapi.io	SearchApi base URL
config.options.engine	string	google	The SERP engine to query
config.options.gl	string	—	Country code
config.options.hl	string	—	Language code

{
  "id": "search-engine_xxxxxxxxx",
  "name": "SearchApi",
  "provider": "searchapi",
  "config": {
    "connection": { "base_url": "https://www.searchapi.io", "token": "${vault://local/SEARCHAPI_KEY}", "timeout": 30000 },
    "options": { "engine": "google" }
  },
  "kind": "ai-gateway.extensions.cloud-apim.com/SearchEngine"
}

DuckDuckGo

GET https://api.duckduckgo.com/?format=json. No API key.

note

DuckDuckGo only exposes the free Instant Answer API, not a full web result list. Results are mapped on a best-effort basis from the abstract and related topics. For full web search, prefer Staan.ai, Tavily, Brave, SearXNG, Google or SearchApi.

{
  "id": "search-engine_xxxxxxxxx",
  "name": "DuckDuckGo",
  "provider": "duckduckgo",
  "config": {
    "connection": { "base_url": "https://api.duckduckgo.com", "token": "", "timeout": 30000 },
    "options": {}
  },
  "kind": "ai-gateway.extensions.cloud-apim.com/SearchEngine"
}

Exa

AI / neural web search. POST https://api.exa.ai/search, authenticated with the x-api-key header (set through connection.token). Exa returns relevance highlights that are mapped to the normalized snippet.

Field	Type	Default	Description
config.connection.base_url	string	https://api.exa.ai	Exa base URL
config.options.type	string	auto	Search mode (auto, fast, instant, deep, …)
config.options.category	string	—	Focus area (e.g. news, research paper, company, people)
config.options.num_results	number	10	Number of results (1–100)
config.options.highlights	boolean	true	Return relevance highlights used as the result snippet
config.options.text	boolean	false	Also request the full page text (used as a fallback snippet)

{
  "id": "search-engine_xxxxxxxxx",
  "name": "Exa",
  "provider": "exa",
  "config": {
    "connection": { "base_url": "https://api.exa.ai", "token": "${vault://local/EXA_API_KEY}", "timeout": 30000 },
    "options": { "type": "auto", "highlights": true, "num_results": 10 }
  },
  "kind": "ai-gateway.extensions.cloud-apim.com/SearchEngine"
}

RAG knowledge base

Instead of the public web, the rag provider searches your own data: it embeds the query with a referenced embedding model, then runs a vector similarity search against a referenced embedding store (your vector database). The most relevant passages are returned in the same normalized result shape, so a RAG knowledge base plugs into providers, agents, the HTTP plugin and workflows exactly like a web engine.

This makes Retrieval-Augmented Generation a first-class, reusable entity: populate an embedding store (for example with the vector_store_add workflow function), point a rag Search Engine at it, and attach it to an LLM provider or agent as a tool so the model can retrieve from your knowledge base on demand.

Unlike the web providers, the rag provider has no connection block. It is configured with two references and a couple of search options:

Field	Type	Default	Description
config.embedding_store	string	—	Id of the embedding store (vector database) to search (required)
config.embedding_model	string	—	Id of the embedding model used to embed the query (required)
config.options.max_results	number	5	Maximum number of passages to return. Overridable per request via max_results.
config.options.min_score	number	0.5	Minimum similarity score (0–1); passages below it are dropped. Overridable per request via min_score.

{
  "id": "search-engine_xxxxxxxxx",
  "name": "Product docs",
  "description": "Search the internal product documentation",
  "provider": "rag",
  "config": {
    "embedding_store": "embedding-store_xxxxxxxxx",
    "embedding_model": "embedding-model_xxxxxxxxx",
    "options": { "max_results": 5, "min_score": 0.5 }
  },
  "kind": "ai-gateway.extensions.cloud-apim.com/SearchEngine"
}

The embedding model and store are referenced by id and are never exposed to the LLM — the model only ever sees a single search tool named after the Search Engine, just like a web engine.

The results map onto the normalized shape as follows:

Normalized field	Value
provider	rag
results[].title	The matched document / chunk id from the store
results[].url	Empty (a knowledge-base passage has no URL)
results[].snippet	The stored passage text
results[].score	The vector similarity score

tip

Give the Search Engine a clear description (e.g. "Search the internal product documentation"). When the engine is used as an LLM tool, that description becomes the tool description the model reads to decide when to query your knowledge base. With no description, a generic "Search the knowledge base … using semantic similarity" prompt is generated for you.