Sam Green, Applied AI team lead, and Aaron Linsky, head of product engineering at Bridgewater Associates, present their research platform in this 17-minute AI Engineer talk.

Make documents fit for research

  • The Vault holds more than 100 million documents from 135 external sources, alongside internal memos, meetings and debate, according to the speakers.
  • Documents are cleaned, governed and tagged to an ontology: the firm’s shared categories for countries, markets, concepts and time periods.
  • Time matters twice: when a document became available and which period its contents describe.
  • Their example asks how AI infrastructure investment affects the Korean won. The Vault-connected agent uses seven sources, compared with one or two in the open-web answer.
  • People need a short list of relevant documents. Agents can consume more material and generate far more searches.
  • Linsky names five design dimensions: recall, precision, filtering, access permissions and scale.
  • The team hand-labeled about 200 examples in less than a week, including plausible-looking but irrelevant documents.
  • Their main metric was precision at ten: how many of the first ten results were relevant.
  • Linsky reports improvement from roughly 60% to above 90% after search changes. Those are internal retrieval results, not investment returns.

Filtering has to work at agent scale

  • Filtering after a similarity search can discard relevant results because too few matching documents made the initial list.
  • Filtering first can require scoring too many documents, slowing the search.
  • Bridgewater switched to turbopuffer, whose filtering approach Linsky says combines fast retrieval with strong coverage of matching documents.
  • He contrasts roughly 100 human queries per hour with agent workloads exceeding 100 per second. Permissions still need enforcement across both.

Move existing habits into the new tool

  • Better search alone failed to change the researchers’ daily routine.
  • Broker-email subscriptions became saved searches; folders became saved collections.
  • Once those habits moved over, users could follow topic feeds, ask questions across documents and hand collections to agents.
  • The platform reached chat, mobile, Slack, development tools and a command-line interface.

Close the research loop

  • An investor workbench proactively proposes topics with a summary, underlying evidence and links to deeper research.
  • Investors decide whether to investigate or ignore a suggestion.
  • The resulting analysis returns to the Vault, making the next research cycle start with what the firm has already learned.
  • The speakers also describe research agents, macroeconomic forecasts and investment signals. They provide no audited return or broad accuracy study.

Caveat: this is Bridgewater’s account of its own platform. The search benchmark and selected demo do not establish better investment performance.

“It’s the same habits, just a new home.” — Aaron Linsky, 12:27