How Bridgewater Turned 100 Million Documents Into an Edge - Sam Green & Aaron Linsky, Bridgewater
Bridgewater Associates' Applied AI and product engineering leads explain how they built the Vault, a governed repository of more than 100 million unstructured documents, and why humans and agents need different things from the same search system. They cover a hand-built benchmark that lifted precision at 10 from about 60% to over 90%, the filtering and scaling choices behind it, and the shift toward proactive agents that bring investors the topics most deserving of their attention.
Bridgewater’s investment process has always depended on data that is carefully built, governed, and aligned to the firm’s own world model. For most of its 50-year history that meant time series models over structured data. Language models opened a much wider field of news, research, policy documents, transcripts, and the firm’s internal memos and meeting debates. The team quickly recognized that this unstructured material needed the same rigor before it could support point-in-time research or systematic study. No vendor offered the control they required, so they built the Vault.
The Vault as a data foundation
The Vault holds more than 100 million documents from 135 external sources, alongside the firm’s internal material, and thousands more arrive every day. Everything is cleaned, permissioned, and enriched against the internal ontology with tags for countries, concepts, markets, and, most importantly, the time period a document discusses. Time is the axis that matters most for investing, because a source that is accurate but months old can mislead an analysis.
Why agents need great data and the right connectors
To show the gap, the team asked an enterprise chat assistant how AI infrastructure investment was affecting the Korean won and whether market participants expected it to continue. The agent found 15 relevant-looking sources but relied on only one or two, and it admitted it had not found enough open-web material to establish a market consensus. Its main source was dated July, even though the question was asked in early October. Three months is a long time in AI and investing, so that source was not usable. The same query run against the Vault behaved differently. The connection prompted the agent to scope its search to October 1 to 8, surface 52 candidate documents, and draw on seven of them to produce an answer it was more confident in. The lesson is that agents become impactful only when they have both strong data and the right connectors to reach it.
Designing search for two different readers
Humans need precision and a high signal-to-noise ratio because their reading capacity is limited, roughly 300 words per minute even for a fast reader. Agents are a different kind of reader. They are far less sensitive to precision, they attend to whatever they are given, they read around 30,000 words per minute, and they scale rapidly. Building for both means looking at several dimensions together: recall, precision, filtering, entitlements, and scale. Some matter mainly to people, some mainly to agents, and many to both. The team focused on precision for people, scale for agents, and filtering for both.
Measuring precision and fixing search quality
Once the Vault was populated, recall was no longer the pressing problem. The top complaint was that “Vault search is bad.” Rather than debate it, the team built a benchmark: about 200 questions, each paired with the best documents found by hand, including plausible near-misses. It took less than a week and was not glamorous, but it made the problem measurable. The key metric was precision at 10, meaning how many of the top 10 results are the right documents. The baseline was about 60%. A series of improvements, culminating in agentic search, moved it into the 90 to 95% range. The benchmark then became the way every later change was judged.
Filtering at scale
Almost every production query carried a condition, whether a date, a source, or a specific asset, so advanced metadata filtering quickly became the first major performance issue. Vector databases classically support two approaches. Post-filtering searches the vector index first and then filters by metadata. It is fast, but recall can be low because globally relevant results may not match the filters. Pre-filtering applies the filters first and then computes distances across every match. Recall is strong, but latency can be severe. The team moved to turbopuffer, whose native filtering searches only the nearest clusters that contain matches and scores those. The result is speed close to post-filtering with recall close to pre-filtering.
Scale was the other pressure point. Human usage is modest, around 100 queries an hour across 300 users who read results over minutes. Agents operate several orders of magnitude higher, easily exceeding 100 queries per second, with agent swarms reaching thousands. turbopuffer handled this load as well.
Habits matter as much as tools
Even with the capabilities in place, researchers were not coming to the Vault every morning. Powerful tools do not change behavior simply by being placed on investors’ desks. The team had to migrate existing habits into a new form. Investors were living in email: broker research arrived directly, they maintained subscription lists, and rules sorted everything into folders. That workflow was familiar, and they trusted it. The Vault rebuilt each habit. Everything became searchable, subscription lists became saved searches, and folders became saved collections. Once users were there, new habits became possible: cross-corpus questions, dynamic topic feeds, and handing collections to agents for further action. That was the point where the team saw clear product-market fit. Demand then pushed the Vault into chat, a phone app, Slack, and developer IDEs, and it was made available to agents through a CLI, because agents work well with command-line tools.
What the Vault makes possible
The first AI analyst, the pocket analyst, can complete hours of research in a few minutes for a few dollars, and it relies directly on the Vault. The researcher extends the same idea across days or weeks. The team is also training custom models to reason more intelligently over Vault data and combining them with foundation models to feed systematic decision-making. The forecaster produces AI-generated macro-relevant predictions, and investment signals convert unstructured data into structured, tradable strategies.
Proactive agents and the investor workbench
The real prize, in the team’s view, is turning agents from reactive resources that answer questions into proactive participants that steer high-value professionals toward the areas that most need their attention. The investor workbench is built around that idea. An agent runs continuously, tracks what matters to each investor, and proposes topics. Each proposal arrives as a card that gives the investor what they need to decide whether to dig deeper or ignore it. The card includes an agentic summary for a quick scan, the exact evidence the agent relied on so the investor can judge the recommendation independently, and a direct link to the research output from the analytical tools if the investor chooses to go further. When investor and agent work through the dynamics together, the results of that collaboration are captured back in the Vault automatically.
Compounding what the firm learns
This feedback loop is the central mechanism. The Vault supplies the right evidence and the right internal and external context at the fingertips of investors and agents, and it captures the synthesis they produce and returns it to the corpus for the next cycle. That is what lets the research circle turn faster and better over time. The underlying lesson is that an agent is only as good as the context it has about your world. The open question the team poses is whether your AI knows what is happening in the world, and whether it remembers what your firm has learned about it.