How Bridgewater Turned 100 Million Documents Into an Edge - Sam Green & Aaron Linsky, Bridgewater

News, broker research, policy documents, internal memos, and meeting debates can all inform an investment decision. But making them available to a language model does not automatically make them fit for investment research. Bridgewater built the Vault to turn that material into a first-class asset. Sam Green, team lead for Applied AI, and Aaron Linsky, who runs product engineering, explain the work behind a repository serving both human investors and agents.

The starting point is Bridgewater’s fifty-year effort to encode its understanding of markets and economies so that the firm can stress-test its reasoning over time. If-then rules on a yellow legal pad developed into a continuing research process: observe a changing world, notice dynamics the existing understanding does not explain, and incorporate what has been learned. AIA, the Artificial Investment Associate, extends that effort by putting a machine at the center of the process.

Historically, much of that work used time-series models over structured data. Those inputs needed careful preparation and alignment with Bridgewater’s internal ontology—its organized representation of countries, markets, concepts, and their place in the firm’s world model. Language models opened access to a wider range of material, including conversations that previously remained scattered across memos and meetings. That material needed the same cleaning and governance, both for questions asked at a particular point in time and for systematic study across time. Green says available vendors did not provide the control Bridgewater wanted, so the firm built its own platform.

At the time of the recording, the Vault held more than 100 million documents from 135 external sources, alongside internal material, with thousands of documents arriving each day. Indexing that volume was only part of the job. Documents were enriched with the countries, concepts, markets, and—especially—the time periods they discussed. A document’s subject and temporal context therefore became information the retrieval system could use, rather than something every reader or agent had to infer afresh.

Green tests the difference with a concrete investment question: “How is AI infrastructure investment impacting the Korean won, and do market participants expect this to continue?” The second half matters. An explanation of a possible currency effect is insufficient; the answer also needs evidence of what market participants expect.

An enterprise chat client searching the open web found fifteen relevant-looking resources but used only one or two in its answer. It also acknowledged that it lacked enough documents to establish a market consensus. Inspecting a heavily used note exposed another problem: it dated from July 2026, while Green had captured the answer in early October. For this fast-changing investment question, he judged a roughly three-month-old perspective too stale. Finding something on-topic had not established that it represented the market’s present view.

What changes when the answer can draw on the Vault? The connection helped prompt the agent to choose a specific interval, October 1–8. Within that window, it found fifty-two candidate documents and used seven in its final response. The important relationship is between the time constraint, the larger investment-specific evidence pool, and the sources actually used: the agent had more material from the relevant period with which to address the expectations question.

The example demonstrates a change in retrieval and evidence coverage, rather than a measured improvement in forecasting or investment returns. Green describes the resulting answer as richer and the agent as more confident. His engineering conclusion is narrower and useful: both the data and the connector matter. The corpus supplies relevant perspectives; the connection helps the agent ask for them under an appropriate time constraint.

Linsky begins the product-engineering account with two different readers. Humans need high precision because every irrelevant result consumes scarce reading time. Agents can process much more material, so the system must also support broader retrieval and much higher throughput. His illustrative comparison is about 300 words per minute for a fast human reader versus 30,000 for an agent. These are the workload assumptions behind the design, not a guarantee that an agent will correctly use everything it receives.

Five dimensions organized the work: recall, precision, filtering, entitlements, and scale. Filling the Vault through the efforts of the data and legal teams had made recall less pressing, but users still complained that search was bad. Having a document somewhere in the collection did not mean the right document appeared near the top of a search. Linsky’s response was: “Let’s benchmark it.”

The team collected questions and manually identified the best documents, including documents that looked helpful but were not. Those misleading candidates mattered: the benchmark needed to distinguish genuine answers from plausible-looking distractions. About 200 examples took less than a week to assemble. The primary metric was precision at ten—the fraction of the first ten results judged to be the right documents. The initial score was about 60%. The users were right.

How much did subsequent search work change that result? Linsky presents a sequence of five improvements, culminating in agentic search, and reports an endpoint around 90–95%. He briefly alternates between calling that endpoint recall and precision; precision at ten is the metric he explicitly defines for the benchmark. The defensible comparison is therefore a reported rise from roughly six relevant results in the first ten to roughly nine or more on this hand-built evaluation—not an explanation of how each undisclosed search change contributed.

Production traces revealed that relevance almost always came with a condition: information available as of a date, documents from a particular source, or research about a specific asset. These conditions made filtering central to retrieval and exposed the Vault’s first major performance problem. A fast search over the wrong eligible documents was not enough.

The two conventional approaches Linsky describes put the expensive work in different places:

• Post-filtering: Search the vector index first, then apply metadata conditions to the top results. This is fast, but the globally nearest documents may fail the conditions. Relevant documents within the permitted subset can be missed because they never entered the initial candidate set.
• Pre-filtering: Apply the metadata conditions first, then calculate distances for every matching document. This preserves strong recall over the eligible set, but scoring all those matches can be slow.

Bridgewater tried and then switched to Turbopuffer for a middle approach called native filtering. In Linsky’s description, it looks only in the nearest clusters that contain matches and scores those candidates. The filter therefore participates in deciding where to search, instead of merely discarding results afterward or requiring distance calculations across every eligible document. He reports speed like post-filtering with recall close to pre-filtering.

Where does each approach narrow the work? The comparison below separates the ordering of filtering and scoring. Post-filtering can lose eligible answers before the filter runs; pre-filtering can leave too many distances to calculate. Native filtering uses matching clusters to narrow the scoring work. That is the specific compromise the Vault needed for queries constrained by dates, sources, and assets.

Human and agent demand also differ at the service level. Linsky describes the human workload as about 100 queries an hour across 300 users, with people spending minutes reading returned documents. Agent workloads can exceed 100 queries a second, with swarms reaching thousands. Reading results in seconds lets agents generate another round of retrieval much sooner; a service sized around human pauses would face a very different load.

Linsky says Turbopuffer handled that scale as well as the filtering requirement. Entitlements remained one of the five design dimensions, but this recording does not explain their implementation. Even with the search requirements covered and new workflows possible, a separate problem remained: researchers still were not opening the Vault every morning.

The existing research workflow lived in email. Broker research arrived directly, subscription lists determined what came in, and rules routed messages into folders. Researchers understood that arrangement and trusted it to keep them from missing material. Better search did not replace the daily habit of receiving, sorting, and checking research.

The Vault gave those familiar activities a new home:

• Subscription lists became saved searches. Researchers could preserve a continuing stream of material matching their interests.
• Folders became saved collections. The organizing habit moved into the shared data platform.
• The accumulated research became searchable. Incoming material could participate in retrieval across the corpus rather than remaining only in an email workflow.

Once researchers had a reason to work there, the Vault could introduce different habits: asking questions across all the data, following dynamic topic feeds, and handing collections to agents for further work. This is where Linsky places the arrival of product-market fit. Demand then spread to chat, a phone app, Slack, and the IDE. Agents received access too, including a CLI—“because agents love CLIs.” The shared platform became useful by accommodating different ways of working, rather than expecting humans and agents to use one identical interface.

The resulting data foundation supports several distinct uses. Some accelerate an investor’s research; others turn unstructured information into inputs for systematic decisions. These are parallel applications of the Vault, rather than stages every document must pass through.

• Pocket Analyst: Bridgewater’s first AI analyst relies on the Vault to perform research. Linsky says it can compress hours of research into a few minutes for a few dollars. That is a reported capability, without a task-by-task quality comparison in the presentation.
• The researcher: The next generation being developed extends the same research idea across days or weeks, rather than only a short session.
• The forecaster: Custom models trained to reason over Vault data work with other foundation models in systematic decision-making processes. The first named example produces AI-generated, macro-relevant predictions.
• Investment signals: Bridgewater has also used the Vault to turn unstructured data into structured, tradable strategies. The talk identifies this use without detailing the signal construction.

Green returns to the opening ambition: an agent should help decide where an investor’s attention belongs, rather than wait for every question. The Investor Workbench is the experience Bridgewater is building around that idea. An always-on agent tracks what matters to an investor and proposes topics that may merit attention. The example returns to AI’s effect on currencies, but changes who initiates the investigation: the system now brings the topic to the investor.

What does the investor need to decide whether to investigate or ignore a proposal? The topic card provides a summary for a quick scan, the specific evidence behind the recommendation, and a direct route into research output from the analytical tools. Those elements support different depths of inspection. A summary helps with triage; the underlying evidence lets the investor judge why the recommendation was made; the research connection supports further work if the topic earns attention.

When the investor and agent investigate together, their findings are automatically captured back into the Vault. This is a concrete form of organizational memory: the synthesis becomes part of the corpus available on the next turn. External evidence and internal understanding no longer have to remain separate from the conclusions reached while using them.

The compounding claim rests on that return path. The Vault supplies internal and external evidence to investors and agents; their collaboration adds new synthesis to the same retrievable collection. Future research can therefore draw on what the firm has already learned, as well as on newly arriving documents. Green closes with two connected questions: does the AI understand what is happening in the world, and does it remember what the firm has learned about it? The Vault is Bridgewater’s effort to make both answers improve together.