How Bridgewater Turned 100 Million Documents Into an Edge

Thumbnail for How Bridgewater Turned 100 Million Documents Into an Edge Watch on YouTube
TL;DR
  1. 1

    Bridgewater built the Vault because unstructured data needed the same cleaning, governance, permissions, and internal ontology as its structured research data.

  2. 2

    A single search system has to meet different needs for humans and agents, with precision for people, scale for agents, and filtering for both.

  3. 3

    The Vault becomes more useful when agents proactively direct investors to topics and return the results of their collaboration to the shared corpus.

Summary

Sam Green and Aaron Linsky explain how Bridgewater built the Vault, a repository containing more than 100 million external and internal documents. The system cleans, governs, permissions, indexes, and tags the material with the firm's ontology, including countries, concepts, markets, and time periods. They compare open-web search with Vault search and show why a larger, time-aware corpus gives agents better evidence. The engineering work focused on precision for human readers, scale for agents, filtering, and entitlements. A hand-built benchmark of about 200 questions moved precision at 10 from roughly 60% to 90-95%. Adoption also required moving familiar email habits into saved searches and collections. The Vault now supports research agents, forecasting, investment signals, and an investor workbench that brings topics to investors proactively. When investors work with an agent, the resulting synthesis is captured back into the Vault, allowing later research to build on it.

Key ideas
01:11

Bridgewater treats unstructured data as part of its research process

Bridgewater spent decades building time-series models from structured data that was cleaned, governed, and aligned with its internal ontology. Language models opened access to news, research, transcripts, policy documents, internal memos, meeting transcripts, and debates. Sam Green says this material needed the same treatment before it could support current research questions or systematic study over time. Since existing vendors did not provide enough control, Bridgewater built the Vault as its own repository and platform. The system indexes, organizes, and prepares external and internal documents for human investors and agents.

05:38

The Vault adds context to more than 100 million documents

The Vault contains more than 100 million documents, with thousands added every day from 135 external sources and from inside Bridgewater. The documents are not simply placed in front of investors. They are aligned to the firm's ontology and enriched with countries, concepts, markets, and the time period discussed. That structure changes how an agent answers investment questions. In Green's example about AI infrastructure and the Korean won, open-web search found 15 relevant-looking resources but used only one or two and could not establish market consensus. The Vault narrowed the question to October 1 through October 8, found 52 candidate documents, and used seven sources.

07:01

Humans and agents need different search behavior

Aaron Linsky separates search requirements for people and agents. Humans need high precision and a high signal-to-noise ratio because they have limited reading time. Agents consume information far faster and are less sensitive to precision when the needed evidence is present. Linsky frames the system around recall, precision, filtering, entitlements, and scale. Precision matters especially for people. Scale matters especially for agents. Filtering matters to both. This split affects the design of the Vault because the same corpus has to support an investor reading a small set of documents and an agent processing very large volumes.

08:32

A small hand-built benchmark exposed poor search quality

When the Vault was full, the main complaint was that its search was bad. Linsky's team collected about 200 questions and selected the best documents by hand, including documents that looked useful but were wrong. The work took less than a week and created a way to measure improvements. Precision at 10, meaning the share of the top 10 results that were correct, was about 60%. After the team applied improvements and added agentic search, it reached 90-95%. The benchmark turned a general complaint into a measurable product problem.

09:39

Native filtering balances recall and speed

Production traces showed that searches usually included conditions such as a date, a source, or a specific asset. Post-filtering is fast because it searches the vector index first, but it can miss globally relevant documents that do not appear in the initial results. Pre-filtering has better recall because it filters first, then calculates distances across the matches, but it can be slow. Linsky says Bridgewater switched to TurboPuffer's native filtering approach. It searches the nearest clusters containing matches and scores those documents, giving it the speed of post-filtering with recall close to pre-filtering.

11:38

Adoption required moving old research habits into the Vault

Better search did not automatically change how Bridgewater's researchers worked. Many researchers were used to broker emails, subscribed lists, and rules that sorted material into folders. Linsky's team brought those habits into the Vault. Subscribed lists became saved searches, and folders became saved collections. Once researchers were using the system, they could ask questions across the corpus, follow topic feeds, and hand collections to agents. The Vault was then made available in chat, a phone app, Slack, an IDE, and a command-line interface for agents.

13:15

The Vault supports research, forecasts, and investment signals

The Vault lets Bridgewater move through its research process faster. The pocket analyst can perform hours of research in a few minutes for a few dollars, relying on Vault data. The researcher extends the same idea across days or weeks. Bridgewater also trains custom models to reason over Vault data alongside foundation models. The forecaster produces AI-generated macro-relevant predictions. The firm is building investment signals that turn unstructured data into structured, tradable strategies. These uses depend on the corpus being organized enough for models to reason over it.

14:20

The investor workbench makes agents proactive

Green describes the investor workbench as an always-on experience that tracks what matters to an investor and presents topics that may deserve attention. Each topic includes an agentic summary, the exact evidence behind the suggestion, and a connection to deeper analytical output. An investor can decide to investigate or ignore the topic. When the investor and agent work together, the resulting learning is automatically captured back into the Vault. This gives later users and agents access to the evidence and synthesis produced earlier, allowing the research process to build on its own output.

"The Vault enables compounding because it provides the right evidence, the right extra internal and external information at the fingertips of our investors and agents, and it automatically captures the synthesis that they generate and puts it back in the corpus for the next turn."Sam Green15:39
Who should watch
  • You are building an internal document system and need to decide how search should differ for people and agents.
  • Your retrieval system produces complaints about relevance, but you do not yet have a hand-built benchmark for measuring search quality.
  • You want agents to bring work to professionals proactively and preserve the results of human-agent collaboration for later use.