search

Web search backed by the Databricks Model Serving native web-search tool.

Unlike a scraping client, the search runs inside a model call: we POST the query to the workspace’s serving endpoint with the provider’s web-search tool spec attached, and the model searches the web and writes the answer. runWebSearch:

1. Resolves a web-search-capable model INDEPENDENTLY of the calling agent’s chat model (the agent may run on a model without web search). A pinned model (request or config) is fuzzy-matched; otherwise the configured fallback order (Gemini, then GPT) is walked to the first web-search-capable endpoint that exists. An explicit but unsupported model is a hard error, not a silent fallback.
2. POSTs to the provider’s REST surface (/serving-endpoints/responses for OpenAI, /serving-endpoints/chat/completions for Gemini) with the mapped tool spec, authenticated as the OBO caller.
3. Returns the synthesized answer plus the cited sources, with citations silently filtered through the configured URL allow-list.

Interfaces

• WebSearchContext

Functions

• resolveWebSearchContext
• runWebSearch