Skip to content

Research knowledge from the web

Sometimes there is nothing to upload. The vocabulary you need is what an industry takes for granted, or what a particular organization does and who it serves, and none of it is written down anywhere you can attach to a form. Research asks several AI models to go and find that context on the open web, each one writing a business documentation document of its own, and then merges their work into drafts you review. Nothing enters your knowledge base until you accept it.

Start a research run

Section titled “Start a research run”
  1. Go to Knowledge and open the Research tab.
  2. Click Start research. The New research request dialog opens.
  3. Under What do you want to research?, describe the business domain, organization, or subject. The dialog suggests the shape: “Describe the business domain, organization, or subject to research. The result is business documentation: entities, attributes, relationships, and events.”
  4. Pick a Target language for the documents the models write.
  5. Choose the Scope: organization-wide, or a single project.
  6. Under Models, switch off any model you do not want in this run. Every model in your organization’s lineup starts switched on, and the choice applies to this run only: the next request opens with all of them on again.
  7. Click Start research.

Watching a run

Section titled “Watching a run”

A research run reports two stages. Researching across vendors… carries a running count of how many models have finished, so (2 / 4) means two of four are done. Consolidating vendor documents… follows while their work is merged into drafts. The run list on the Research tab shows the same stage as the run’s status, so you can tell a researching run from a consolidating one without opening it.

Underneath, the event feed names each model twice: once when its web research completes, with the number of sources it cited, and once when it writes its business documentation. Then it records how many distinct sources the lineup cited between them, and reports the consolidation step as it starts and finishes, down to the number of drafts it produced.

When the run finishes, its status becomes Awaiting review and the drafts appear for you to accept or reject individually. The Research tab filters runs by Awaiting review, Finalized, Failed, and Archived.

What a research run actually does

Section titled “What a research run actually does”

Each model researches the web with live search and writes one business documentation document of its own, saved to the run the moment it lands. Once the last model finishes, consolidation runs as a step of its own: the documents are merged into one set of drafts, and the web addresses the models actually cited are attached to those drafts as sources.

The cost is known before you start: one research call per model plus one consolidation call, rather than an amount that only settles as the work runs.

When research is not available

Section titled “When research is not available”

Research runs on a vendor lineup a platform administrator enables. Until one is enabled, Start research stays disabled, and hovering it explains why: “Research is unavailable: no research vendor lineup is enabled”. Ask a platform administrator to enable a lineup.

Review the drafts

Section titled “Review the drafts”

Research never writes to your knowledge base on its own. Each draft arrives with the sources it was built from, so you can check a claim against the page it came from before you accept it.

Work through the drafts with Accept this draft and Reject this draft, or tick several and use Accept selected and Reject selected. Edit (advanced) opens a draft for editing before you decide. Abandon discards the whole proposal. Rejected drafts leave nothing behind.

Citations and provenance

Section titled “Citations and provenance”

Three things on the review screen tell you where a draft came from.

Cited sources lists every web address the lineup cited, each with a badge naming the models that cited it. An address every model reached is corroboration; an address only one model reached is a lead worth checking. They are links, so you can open the page a claim came from before you accept the draft that makes it.

Vendor documents sits below the cited sources and holds one row per model: its name and the length of the business documentation it wrote. Expand a row to read the document in full. These are the unmerged originals the drafts were consolidated from, so when you want to know what a model actually said before consolidation weighed in, its own words are one click away.

Found by sits at the top of each draft, under its name, and names the models that proposed that draft. A draft found by the whole lineup is one every model reached independently. A draft found by one model is that model’s alone, kept because consolidation judged it worth keeping. Neither badge decides anything for you: they show how much agreement sits behind a draft so you know where to spend your review time.

A model that researched the subject without returning a usable web address still contributes its document to consolidation. Drafts built from it name the model in Found by and carry no cited address, because DeltaVault attaches the addresses the models really cited and never invents one to fill the gap.

Section titled “Related guides”