Web Search
AI-Starter supports built-in web search capabilities with select provider models. It is configured through the Platform Metadata settings of an assistant.
Supported Platform Metadata
| Type | Description |
|---|---|
| BedrockNova | Web search with Nova 2 Lite |
| BedrockOpenAIResponse | Web search with OpenAI GPT models on Amazon Bedrock (via bedrock-mantle) |
| OpenAIResponse | Web search via OpenAI's Response API |
| GoogleGemini | Google Search and URL context |
OpenAI
OpenAI's web search feature is available through the following method.
- Web search tool via OpenAI Responses API
Data Handling
- As with standard models, input data is not used for model training.
- We have confirmed with the provider that input data is not used for improving the web search service.
For details, see Web search | OpenAI.
Gemini
Gemini's web search feature uses the Grounding with Google Search capability and is available with the following models.
- Gemini 3.8 Flash
- Gemini 3.8 Flash Thinking
- Gemini 3.5 Flash-Lite
Data Handling
- As with standard models, input data is not used for model training.
- Per the Grounding terms of service, data is retained by Google for 30 days.
- Unless explicitly instructed by the customer, Google does not use retained data for debugging or testing.
Display Requirements
Due to the Grounding with Google Search display requirements, search suggestions and grounding evidence are shown only on the first view. They are not displayed in history or logs.
For details, see the following official documentation.
Nova
Nova's web search feature is available by enabling the web search option on the following models.
- Nova 2 Lite
Caution
Nova's web search feature requires the US East 1 (N. Virginia) region.
For details, see the following official documentation.
GPT on Bedrock
Among the OpenAI GPT models available on Amazon Bedrock, models that use the bedrock-mantle endpoint support web search. Searches use an AWS-hosted tool.
For supported models, see Web search | Amazon Bedrock.
Controlling external web access
With Bedrock Web Search, you can control whether external web access is allowed when a search result page is unavailable in the cache.
AI-Starter controls this setting through Enable External Web Access in the assistant's Platform Metadata. It is disabled when omitted.
When external web access is enabled, Fetch first checks the AWS-hosted cache and retrieves content from the external web when suitable cached content is unavailable. In this case, request data may leave the AWS boundary. Review your data-governance requirements before enabling it.
Note that these models handle data differently from other Bedrock models. Please also review Abuse detection.