Web Search

3 min read

The tool can be activated within the following spaces:

Space configuration

What It Does

The Web Search tool gives AI assistants the ability to search the internet in real time, read web pages, and use the retrieved information to answer user questions with properly cited sources.

When a user asks a question that requires up-to-date or external information, the assistant automatically formulates search queries, retrieves results from one or more search engines, reads the relevant pages, and synthesizes the content into a grounded answer.

Key Advantages

Multiple search engines — Choose from Google, Bing (Grounding with Bing), Brave, Jina, Tavily, Firecrawl, VertexAI (Grounding with Gemini), or connect your own custom API endpoint.
Intelligent query handling — Queries can be automatically refined and optimized before being sent to the search engine, improving result quality.
Three search strategies — V1 performs straightforward keyword searches; V2 lets the AI plan a multi-step research approach with parallel searches and direct URL reads; V3 (Experimental) lets the AI drive an iterative agent loop, alternating cheap snippet searches with on-demand full-page reads.
Configurable web page reading — Multiple readers are available to fetch and convert web pages into readable text, including options for JavaScript-heavy sites.
Content processing pipeline — Retrieved content is automatically cleaned (removing navigation menus, cookie banners, etc.), chunked, and optionally summarized before use.
Relevancy sorting — An AI-based re-ranking step ensures the most relevant content is prioritized.
Hallucination detection — Built-in evaluation checks help detect when the generated answer may not be well-supported by the retrieved sources.
Source citation — Answers include proper references back to the original web sources.
Query review (Elicitation) — An optional feature that lets users review and modify proposed search queries in a UI form before they execute, providing transparency and control over what is searched.

Important Warning — Please Read Carefully

The query generated for the web search tool automatically uses surrounding context from the current conversation.

This means that any information shared previously in your conversation — including potentially sensitive or personal data — may be included in the search query sent to external search engines.

If your discussion contains confidential, private, or sensitive information, you should avoid using the web search tool, as this process can potentially lead to unintentional data exposure or leakage.

For platforms with stricter requirements, the optional Argument Screening experimental feature adds an LLM-based pre-screen that can block searches when sensitive information is detected; see the Experimental Features page.

Search Modes

Mode	Status	Description
V1 — Basic Search	Stable	The orchestrator AI sends a single search query. The tool optionally refines it, runs the search, reads the pages, and returns processed content. Simple and fast.
V2 — AI-Planned Research	Stable	The orchestrator AI creates a structured research plan with multiple steps (searches and direct URL reads). All steps execute in parallel for faster results. Better for complex or multi-faceted questions.
V3 — Agent-Driven Research	Experimental	The orchestrator AI drives the loop itself: each tool invocation runs a single command — either a cheap snippet-only `search` or an on-demand full-page `read_urls` — and the model decides what to do next based on what it found. Best for open-ended research where the right next step depends on intermediate results.

Main Configuration Layers

Layer	What It Controls
Search Mode	Whether to use V1 (basic), V2 (AI-planned), or V3 (Experimental, agent-driven).
Search Engine	Which search provider to use (Google, Bing, Brave, etc.) and how many results to fetch.
Web Page Reader	How web pages are fetched and converted to text. Choose between built-in readers or API-based options.
Content Processing	How fetched content is cleaned, chunked, and prepared. Includes options for AI-based summarization and privacy filtering.
Answer Quality	AI-based relevancy sorting of results and hallucination detection on the generated answer.
Source Citation	Instructions for how the AI should cite web sources in its responses.
Experimental	Query Review (elicitation), Argument Screening (PII pre-screen), Tool Response System Reminder.

Where to find what

Detail pages are split by audience so each reader can go straight to what is relevant to them:

Technical — for space administrators configuring the tool inside a Space. Each page documents the settings as they appear in the Spaces UI: Search Mode Configuration, Search Engine Configuration, Web Page Reader Configuration, Content Processing Configuration, Quality and Evaluation, Experimental Features.
Platform / Architecture — for security and architecture teams reviewing how the tool makes external calls, what data leaves the platform, and how each search mode and engine is wired internally. Sequence diagrams included.
Platform / Infrastructure — for DevOps and platform engineers deploying or operating assistants-core. Reference-style content on activation, credentials, identity, proxy egress, feature flags, and environment variables.
Security Risk Assessment — security review covering infinite-loop, prompt-injection, internal-KB-isolation, and recursive-crawling risks. See Security Risk Assessment.