Tools: research
UserBehind each of the two research chips in chat — Research (research mode) and Web — sits one built-in tool. This page only shows the tool view — how the mode works as a feature is described in Research mode.
The two tools
Section titled “The two tools”| Tool | Purpose | Kind | Available |
|---|---|---|---|
Research the knowledge base (research_vault) | Enumerates the entire accessible knowledge base and produces a report from it. | Read | only to the research agent |
Research the web (research_web) | Researches a question on the web with OpenAI’s web search and returns a report with numbered sources. | Read | to the web research agent, not to chat |
How a run starts — no model in the chat
Section titled “How a run starts — no model in the chat”The run is not started by a tool of the chat assistant but by the chat itself: if one of the chips Research or Web is on when you send, the chat starts the matching seeded agent in the background and ends the turn at once — without any model call. That’s why you first get only a placeholder notice and the actual report minutes later.
research_vault — the map-reduce engine
Section titled “research_vault — the map-reduce engine”research_vault is not available to the chat assistant. It is the sole tool of the seeded research agent, which the Research chip launches. It sifts the knowledge base in full — not via top-k search — evaluates the passages selected that way with one model call per matched document, and hierarchically condenses the individual findings into a Markdown report. → Full enumeration instead of top-k
research_web — web research
Section titled “research_web — web research”research_web is not available to the chat assistant either. It is the sole tool of the seeded Web Research agent, which the Web chip launches. It hands the question to OpenAI (USA) in one background call; OpenAI researches it with its own web search and writes a report with numbered, linked sources. → Web research · Research or Web?
When a run starts at all
Section titled “When a run starts at all”Both kinds of research are gated twice. A run only starts when both conditions hold:
- The administrator has enabled that research — for Research the deployment killswitch
Research:ChatToolEnabled, for Web the Allow web research switch with a usable model in system settings (both off by default). - You have switched on the matching chip for this very message.
If no chip is on, chat answers normally via RAG. Note that condition 2 resets itself with every message sent: the chip is read and cleared on send. Each user also only ever has one research run at a time. → Enabling research mode
Cost & runtime
Section titled “Cost & runtime”Both tools only read, so they change nothing in your data. Starting a run costs nothing, because no model is involved. The run itself is a different matter:
- Runtime: minutes — usually a few, with many matches considerably more. Aborted after the maximum run time set on the research agent, 20 minutes by default.
- Cost: a multiple of a normal RAG answer. Sifting every accessible document, however, runs locally and without a model call; only the selected passages go to the model, one call per matched document. Consumption therefore follows the number of matches for the question, not the size of the knowledge base.
- Web: usually 5 to 20 minutes, aborted after the maximum run time of the web research agent (40 minutes by default). You pay for the tokens of the web research model; OpenAI charges each web search on top. The number of web searches is shown at the end of the report.
That’s why the mode is switchable per message rather than permanently on.
Permissions
Section titled “Permissions”The run is RBAC-scoped: it only enumerates the vaults and documents the triggering user is allowed to see. So a research run cannot surface anything that would stay hidden in normal chat. Web research does not read your knowledge base at all; only your question goes to OpenAI. → RBAC & confidentiality