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

Sam Green, Bridgewater Associates · Aaron Linsky, Bridgewater Associates

[Talk page](https://ai.engineer/talks/CLve4Nx_RwI-bridgewater-turned-100-million-documents-into-edge) · [Video](https://www.youtube.com/watch?v=CLve4Nx_RwI)

## How Bridgewater Searches 100 Million Documents for Investment Research

Sam Green and Aaron Linsky explain how Bridgewater’s Vault combines governed investment data, search built for humans and agents, and familiar research workflows—then captures what investors learn for the next question.

![The Vault diagram connects gathering evidence, reasoning and learning, with new synthesis captured back into the shared corpus.](https://ai.engineer/api/talk-essay-media/9910d017ae75a33a10f7355b89a3f1885ca852e895ff2541f0823ee19d4c287c)

The Vault diagram connects gathering evidence, reasoning and learning, with new synthesis captured back into the shared corpus.

### Give unstructured data the same rigor as market data

[0:12](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=12s)

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 slide lists about 100 million documents across the Vault and 135 external sources. A diagram labels concept, country and market alongside a time arrow.](https://ai.engineer/api/talk-essay-media/24055b50094e99af1f5b66398866a6b304b7ce688c06dac04fda573b140c265f)

The Vault combines Bridgewater’s internal material with data from 135 external sources. Its roughly 100 million documents are organized by concept, country, market and the time period discussed.

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 in total, combining material from 135 external sources with Bridgewater’s own records. Thousands of documents arrived 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.

### The same Korean won question, two evidence sets

[4:14](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=254s)

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 Vault-connected answer uses an October 1–8 date window, 52 candidate items and seven cited sources for the Korean won question.](https://ai.engineer/api/talk-essay-media/d7b74ceb4de35cdd06940c9375ec93dd0112929c8db2295778643650b019cf77)

The Vault-connected answer uses an October 1–8 date window, 52 candidate items and seven cited sources for the Korean won 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.

### Turn “Vault search is bad” into a benchmark

[6:54](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=414s)

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, manually identified the best documents for each, and marked 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.

The benchmark slide lists five changes: fixes to document ingestion and parsing, adjustments to reciprocal rank fusion weights, a reranker, query expansion, and agentic search. By the time the team added agentic search, Linsky reports a score around 90–95%. He calls that endpoint “recall at ten or precision at ten,” having defined the 60% baseline as precision at ten. He gives no separate scores for the five changes.

![The search benchmark slide describes about 200 manually judged examples and lists successive retrieval improvements.](https://ai.engineer/api/talk-essay-media/2a75fb7a6d121888bf2b312a8965a1c5ff2e4f7d7bc3ca814e947cd2678fddc7)

The search benchmark slide describes about 200 manually judged examples and lists successive retrieval improvements.

### Filtering changes which nearest neighbors matter

[9:31](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=571s)

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.

For an investor asking what was known as of a particular date, that meant the cutoff could guide which clusters the system searched.

Three ways to combine similarity search and metadata conditions
Read each branch independently. The location of the filter changes both which documents can survive and how many documents need scoring.
- Search vector index — Find globally nearest candidates before applying metadata conditions.
- Filter top results — Fast, but candidates can fail the conditions and eligible answers can be missed.
- Find all metadata matches — Restrict the eligible documents before distance calculations.
- Score every match — Strong recall, but calculating distances for all matches can be slow.
- Find nearby clusters containing matches — Use metadata matches when choosing which nearby clusters to inspect.
- Score candidates there — Reported to approach post-filtering speed with recall close to pre-filtering.
- Search vector index → Filter top results (Apply conditions after candidate retrieval)
- Find all metadata matches → Score every match (Calculate distances for all eligible documents)
- Find nearby clusters containing matches → Score candidates there (Limit scoring to candidates in selected clusters)

### Agents change the query rate

[10:53](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=653s)

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.

![The workload slide contrasts about 100 human queries per hour with more than 100 agent queries per second.](https://ai.engineer/api/talk-essay-media/c61b6f4883f0c0d1d742158a390f209f06ebb93de67882f7b8a0a9f83641cbd1)

The workload slide contrasts about 100 human queries per hour with more than 100 agent queries per second.

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.

### Move the email habit before asking for a new one

[11:23](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=683s)

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 workflow slide maps email research, subscription lists and folders to Vault search, saved searches and collections.](https://ai.engineer/api/talk-essay-media/11ff006f6b80db963a780d2b72f883a8df3984c0f75e0a6d4820768e8982d99a)

The workflow slide maps email research, subscription lists and folders to Vault search, saved searches and collections.

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.

### From retrieved documents to research, forecasts, and signals

[12:53](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=773s)

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.

### Propose what deserves attention—and remember what follows

[14:05](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=845s)

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.

![The investor workbench pairs an agent summary with supporting evidence and a route into deeper research.](https://ai.engineer/api/talk-essay-media/ad487484084e409d6fd469b53ca1aeb5034930bb752918838c571819e1a63c09)

The investor workbench pairs an agent summary with supporting evidence and a route into deeper research.

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.

## Key takeaways

- Investment retrieval needs governed, time-aware data. In the Korean won example, the Vault connection helped constrain retrieval to October 1–8 and expanded the answer’s evidence from one or two used sources to seven.
- A hand-built benchmark of about 200 examples turned a search complaint into a measurable problem. Initial precision at ten was about 60%. After adding agentic search, Linsky reported around 90–95%, calling the endpoint “recall at ten or precision at ten.”
- Metadata conditions belong in the retrieval design. Post-filtering risks losing eligible answers, pre-filtering can require too much scoring, and the native-filtering approach described in the talk narrows scoring through nearby clusters containing matches.
- Human adoption required migrating a trusted email workflow: subscription lists became saved searches, folders became collections, and those collections could then support agent work.
- The Investor Workbench connects proactive topic selection with inspectable evidence and deeper research. Capturing the resulting human-agent synthesis back into the Vault makes it available to future investigations.

## Related talks

- [From Raw Documents to AI-Ready Data — Leo Platzer & Jeff Koss](https://ai.engineer/talks/wzWNYDY7toc-from-raw-documents-ai-ready-data-leo) — 
- [Beyond RAG: Build a Relational Context Engine from Scratch — Peter Werry, Unblocked](https://ai.engineer/talks/spxoDAfh5ig-beyond-rag-build-relational-context-engine) — 
- [The Investor's Edge Lives in Their Head — Anekha Sokhal, Moshi](https://ai.engineer/talks/UHbODjuviAU-investors-edge-lives-in-their-head-anekha) — 

## Slides

### Unstructured data as an asset

[0:32](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=32s) [Three principles sit beneath the Vault’s claim about the value of unstructured data.](https://assets.ai.engineer/sites/aiecode2025/sha256/0d259193caee6bad3b81ffb24991c1f63a5c396d9a0fc5575d23948966168608.webp)

The Vault turns unstructured data into a first-class asset.

### How Bridgewater studies markets

[2:05](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=125s) [A handwritten note and two process diagrams trace the approach to understanding markets and economies.](https://assets.ai.engineer/sites/aiecode2025/sha256/fdd9ae9d9a9ad197f0e860f6c6724147212a617f67b92581975ca982e377512e.webp)

Pursuing the best understanding of how markets and economies work.

### From tabular to unstructured data

[2:56](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=176s) [Two panels contrast time-series models with language models working on unstructured content.](https://assets.ai.engineer/sites/aiecode2025/sha256/8b736787a067d1a8f3172c3eeb5814ef641fe004e322753c3d0aea6c7516c98f.webp)

Time series models on structured data · Language models with unstructured content

### What goes into the Vault

[3:25](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=205s) [A diagram groups indexed data by concept, country and market beside the repository size and source count.](https://assets.ai.engineer/sites/aiecode2025/sha256/98ed29756eb2895cb6d5cf9cc41cac639f7243f4db8ecc4b0a1230e5993763c4.webp)

So we built the Vault. ~100M documents · 135 external sources

### Where open-web search falls short

[5:29](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=329s) [A sample answer sits beside three limitations, including sparse and outdated sourcing.](https://assets.ai.engineer/sites/aiecode2025/sha256/f1f3518c58f1462770724a16bb919d829b18fb82eb1f2bfa6d7036feb66a5adb.webp)

Standard tool: out of date, under-sourced

### Adding time to the query

[6:23](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=383s) [Green highlights in the Vault response indicate its recency reasoning and use of sources.](https://assets.ai.engineer/sites/aiecode2025/sha256/a499967c5f4f7d9f8ccb3c2b7a4e47356b109c4d14247608b669379a9848c51c.webp)

The Vault: Searching with Time as Part of the Query

### Humans and agents consume search differently

[7:55](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=475s) [Two columns compare precision, recall, scale and reading speed across the two audiences.](https://assets.ai.engineer/sites/aiecode2025/sha256/e5243007a5304c8e792b86c1ed173225a8886a81d995574d4d557e3404859cce.webp)

Two very different consumers

### Benchmarking Vault search

[8:56](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=536s) [A feedback complaint leads into a manually judged benchmark assembled from real questions.](https://assets.ai.engineer/sites/aiecode2025/sha256/a3e5236f6c23af2bbfd9d1fea48e8411db3b32040625bfeb6ae91fe6e7368be6.webp)

#1 Complaint: “Vault search is bad”

### Filtering the search set

[9:47](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=587s) [Date and topic filters narrow the document set before relevance ranking.](https://assets.ai.engineer/sites/aiecode2025/sha256/325e3e47e24bf826fe2a4a7d9154ac56a3e44dce5aac0e96188ea2b3c1e22439.webp)

We want the most relevant documents.

### Postfiltering, prefiltering and native filtering

[10:48](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=648s) [Three grids show the recall and speed tradeoffs of alternative filtering strategies.](https://assets.ai.engineer/sites/aiecode2025/sha256/1d695f3d2e4ba6f50fe6be99658ff66a4f08fd70b30e9f8657fc8e2fd58f9bee.webp)

Postfiltering v Prefiltering

### Serving agent-scale queries

[11:06](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=666s) [The human-scale and agent-scale columns contrast requests per hour with requests per second.](https://assets.ai.engineer/sites/aiecode2025/sha256/5eaa10e7919b60c5a3c56745a94465bb0559ab5b141ab0aba7303e3ae9b68b18.webp)

100 Queries per Hour · 100+ Queries per Second

### From familiar searches to new workflows

[12:29](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=749s) [Three columns map past habits through Vault features to new ways of working.](https://assets.ai.engineer/sites/aiecode2025/sha256/767f8596684e6e5a5dde7035382b82141f6af794455f6139fff2ae84357562ec.webp)

Same habits, new home → new habits

### Access from the tools people already use

[12:56](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=776s) [Browser, chat, memo, Slack and IDE entry points sit beside API, CLI and MCP access for agents.](https://assets.ai.engineer/sites/aiecode2025/sha256/ffca863c99fb0a6d6d39a7357c790a046cdcfbbf18c5e826ffc129a506a4811c.webp)

Wherever the work happens.

### What agent access makes possible

[13:50](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=830s) [Pocket Analyst and Researcher feed custom reinforcement-learning models and downstream decision tools.](https://assets.ai.engineer/sites/aiecode2025/sha256/23cb3429c9c09a00e59b25db5b4149e806dd6a126085658b3145daa9acf305cd.webp)

What agent access makes possible.

### The Investor Workbench

[14:26](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=866s) [A board of items under “What am I missing?” illustrates a shift from answering questions to steering attention.](https://assets.ai.engineer/sites/aiecode2025/sha256/d6774fbdb0cbb2810defe645aa422de2578b1e88df8b116cf0c3764ca7638d6b.webp)

Building the Investor Workbench

### Evidence alongside the summary

[15:13](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=913s) [A workbench example pairs a written summary with a reading list and highlighted evidence.](https://assets.ai.engineer/sites/aiecode2025/sha256/4a66662712a620e597d311dccc7a7289c8e05fd3a946c327281f4170720b3e70.webp)

Summary and evidence, ready for action.

### Capturing collaboration in the Vault

[15:28](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=928s) [An interface diagram shows where the work of humans and agents is recorded.](https://assets.ai.engineer/sites/aiecode2025/sha256/6c7967a4e6b07c7066df6bb8b2190cc69b88d5b289187398419421a0fab080e6.webp)

The Vault captures the human + agent collaboration.

## Full transcript

[0:12](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=12s) Good morning, everyone. Uh, I'm Sam Green, um, team lead for Applied AI at Bridgewater Associates, and this is my colleague Aaron Linsky. And today we're here to talk to you about the Vault, the data platform that we've built at Bridgewater to turn unstructured data into a first-class asset in our investment process. And as part of the talk this morning, we'd like to share three lessons that we've learned building the Vault. The first is that humans and agents have very different needs from a data

[0:41](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=41s) platform, and so you have to build for those different needs separately. The second is that the greatest impact occurs when a single tool can meet both of these sets of needs at the same time. And the third is that the real prize is when you can build the foundations needed to make agents a proactive participant in your investment process or whatever you're doing with your, with your, your tools, uh, instead of a reactive, uh, reactive tool for your humans to use. Uh,

[1:12](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=72s) before I, I dive deeper into talking about the Vault, I wanted to start with some, some background about Bridgewater. Over the last fifty years, Bridgewater has been in the constant pursuit of the world's best understanding of how markets and economies work and turning that into a systematic encoding that we can use to stress test our logic over time. And what originally began as a set of if-then rules on a yellow legal pad evolved into a continuous compounding process with humans at the center constantly

[1:41](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=101s) looking for the next thing that's changing in markets. And more recently, we've been working on building AIA, the Artificial Investment Associate, which puts a machine at the center of this continuous process. Across all these phases of evolution, we've had a constant philosophy in mind, which is that we want to be constantly looking at how the world is changing and what's going on in the world, noticing what, what dynamics are not yet encoded in our understanding, and use that learning to improve and

[2:11](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=131s) compound for the future.

[2:14](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=134s) For most of Bridgewater's history, that meant building and overseeing time series models on structured data, and that structured data needed to be carefully built so that it could be safe for use in our research process and aligned to our internal ontology and world model so that we could get great leverage out of using the, the data that we had built. With the emergence of language models, we had a huge new opportunity to engage with a, with a much wider world of data, um, you know, news and research,

[2:44](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=164s) transcripts, policy documents, and all of the internal conversation that was previously just in the ether, so the internal memos and meeting transcripts and debate happening inside the firm. We quickly realized that all of that data needed the exact same treatment to be fit for purpose. It needed to be carefully cleaned, governed, and made ready for use in both point-in-time research questions today and also our systematic study over time. And when we looked out at

[3:14](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=194s) the landscape of data vendors and data platform providers, there was nothing that provided us with all of the control that we needed to build and maintain our data advantage. And so we built the Vault, which is our repository of external and internal unstructured data, um, and the platform that indexes, organizes, and makes all of it ready for use by our human investors and agents. Today, the Vault has more than a hundred million documents in it, with

[3:44](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=224s) thousands more added every day, uh, minute by minute, from a hundred and thirty-five external sources, and again, growing all the time, as well as all of the data from inside of Bridgewater. And crucially, all that data is not just dumped on our investors' desks. It is carefully aligned to our internal, internal ontology with, uh, enrichments like the countries, concepts, markets, and most importantly, time period discussed in the document. To bring

[4:14](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=254s) this to life, I want to start with a, a quick motivating example. So what I did here is I asked a fairly basic representative investment question to an enterprise chat client. I asked, "How is AI infrastructure investment impacting the Korean won, and do market participants expect this to continue?" And as, as you're all I'm sure familiar with, um, the agent goes to the web and looks for resources it can draw on because I've explicitly asked it to go find, uh,

[4:44](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=284s) other people's perspectives on the question. And you can see it, it zooms in and it finds fifteen relevant-looking resources to draw on, uh, before it gives me back its answer. And when I, when I zoomed in on the answer it gave back, a couple of things stood out to me. The first is that even though it found fifteen relevant-looking sources, it only picked one or two to answer, to answer my question. And the second is down at the end of this response, where the agent itself acknowledges that it hasn't answered my

[5:14](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=314s) question. It, it didn't find enough documents on the open web to establish a market consensus and meet the, meet the parameters that I laid out. When I double-clicked even further, the problem got even worse. One of the notes that it primarily drew on to answer my question was published way back in July of twenty twenty-six, even though I asked this question and captured this answer just at the end of last week in early October. In the age of AI and in

[5:44](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=344s) investing, for-- three months is a massive amount of time. It's not even close to good enough to draw, uh, draw on this resource to answer my question. When I ask this exact same query attached to the Vault, something totally different happens. The first is that the connection to the Vault helps to prompt the agent that it should look for a specific period of time when answering my question, so it zooms in on October first to October eighth. It finds a much wider array of resources, fifty-two candidate

[6:14](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=374s) documents that are relevant to the query over this spec-specific period of time because of the Vault's much more significant corpus of content that's relevant to investing questions. And when it ultimately answers my question, it draws on seven instead of only one or two and gives me back a rich answer that the agent is co-- more confident is going to address my question. And so for us, this is the, the big lesson of-- from building the Vault, uh, to kind of motivate the talk today, which is that it takes great

[6:43](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=403s) data and the right connectors to make agents impactful. And I'm gonna talk-- hand over to Aaron now to talk about some of the product engineering takeaways from building the Vault.

[6:54](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=414s) Thanks, Sam. So hi, I'm Aaron. I run product engineering at Bridgewater. I'm gonna take us through a few of our learnings from building the Vault over the past few years. I'm gonna start with one that might seem obvious, but I still think it's worth saying, which is search is hard.

[7:11](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=431s) But it gets even harder if we're designing for two different consumers. In, in this case, we're designing for humans and agents. I'm gonna go over the difference here, starting with us. So humans need precision and a high signal-to-noise ratio because we can only read so much. Even a fast reader manages about three hundred words a minute, and personally, I'm much slower. Agents are a different kind of reader. They're far less sensitive to precision. If the data is there, they'll find it, and

[7:41](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=461s) basically anything you hand to them will get attended to. They scale super quickly, and they come-- and they consume about, uh, a hundred times faster, thirty thousand words per minute. And so we need to build Vault for both of these. And when we're thinking about the dimensions that matter, there are these five on, on the board: recall, precision, filtering, entitlements, and scale. Some of them matter to people, some matter more to agents, many to both.

[8:12](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=492s) I'm gonna go deep on three of these: precision for people, scale for agents, and then filtering for both. So precision. Once we've filled the Vault, our-- which was a heroic effort by our data and legal teams, recall was no longer a pres-- a pressing issue. However, the number one complaint was Vault search is bad. As the product owner, it's painful to hear that, but my response was, "Let's benchmark it." And that's what we did.

[8:42](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=522s) We collected questions and found the best documents by hand, including ones that looked helpful but weren't. In all-- in total, it was about two hundred examples, and it took us less than a week. It was not fun or very glamorous, but it was very helpful. And after we ran it, it was very clear search was bad. The key metric we were looking at is precision at ten, which is of the top ten results, how many are the right documents? And we were hitting about sixty

[9:11](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=551s) percent. So the people were right, it was bad. But now we had a way to measure all of the improvements, and here on the board are five of them. And by the time we got done with agentic search, we were up in the ninety, ninety-five percent, uh, recall at ten or precision at ten.

[9:31](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=571s) All right, so that was precision. Now on to filtering. Like anyone, we want the most relevant documents. But when we looked at the production traces, what we were seeing is that almost always there was some condition, either as of a certain date or from a certain source or about a specific asset. And all of that requires advanced filtering, and that's where we ran into our first major performance issue. So I'm gonna quickly go through this of, uh, a

[10:01](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=601s) little bit technical here of there's classically two ways to fil-filter a, a vector database. Post-filtering will use that vector index first and then filter the top results by metadata. It's fast, but recall could be pretty low because the globally relevant results might not match your filters. On the other side, we have pre-filtering, where we'll filter first and then calculate distances for every one of those matches. The advantage there is that the recall is very good, but it can be very slow, and

[10:31](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=631s) we needed something in between. So that's what led us to try and then ultimately switch to Turbopuffer. Their approach, which they call na-native filtering, looks only in the nearest clusters that contain matches and then scores just those. What you end up getting is something that's fast like post-filtering, but with close to the same recall as pre-filtering.

[10:53](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=653s) And lastly, I'm gonna talk about scale. Human scale for us is pretty pedestrian. A hundred queries an hour. We have three hundred users. They're gonna search, you know, every so often. It takes our users minutes to read the docs that they're, that they're receiving. Agents, on the other hand, are a few orders of magnitude beyond that. Easily a hundred plus queries a second, and agent swarms can hit thousands easily. And they obviously will read, as I said earlier, will read docs in seconds. So we needed something that could

[11:23](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=683s) scale as well, and again, Turbopuffer handled that easily too. So we were covering all five dimensions. I didn't talk about entitlements, but I'll be talking about that later today. Um, but we covered it all. It unlocked new workflows, but it didn't change how our researchers worked. They still weren't coming to Vault every morning. And that's my second lesson that I wanna share here, is that no matter how amazing the tools are, we can't just drop them on our investors' desks and expect them to be used. We need

[11:53](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=713s) to migrate their old habits into a, a, a new form. And so we looked at what they were already doing. Many of them were living in email. The research came straight from the brokers. They had subscribe li-re-- lists. They had rules that filtered everything into folders. It worked, it was familiar, and they knew that they weren't missing anything. So we had to build all of these features into the Vault. So that's exactly what we did. Everything's searchable, the subscribed lists

[12:23](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=743s) became saved searches, and folders became saved collections. It's the same habits, just a new home. And once we had them there, we can open up new habits. So we can ask-- they can ask questions across all the data, as Sam showed. They can follow dynamic feeds by topic, and they can hand those collections over to their agents to take further action. And that's when we clearly had product-market fit. And once we had that fit, the users wanted it everywhere, so we put it in chat, we made a phone app,

[12:53](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=773s) it's in Slack in their IDE, and of course, we availed all of that to the agents as well, um, including a, a CLI because agents love CLIs. And so the main question here is, okay, so what does this enable? Sam's gonna talk about what we're building right now, but first and foremost, it's helping us to move around that research circle faster. With our Pocket Analyst, which is our first AI analyst, it can do hund-- it can do hours-- it can do hours of research in just a few minutes for a

[13:23](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=803s) few dollars, and that's relying on the Vault. And then we're expanding that now with the researcher, which is the next generation. It's the same idea stretched across days or weeks. We're also able to train custom models to more intelligently reason over the Vault's data and use those together with other foundation models to feed systematic decision-making processes. The forecaster is our first one, which produces AI-generated macro relevant predictions. And then we

[13:52](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=832s) als-also have, uh, built investment signals from the Vault, turning unstructured data into structured tradable strategies. Now I'll hand it to Sam to take us through Vault's latest capability.

[14:05](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=845s) Thanks. So like I alluded to at the beginning of the talk, the real prize is when you can take agents from a reactive resource that can be ready to answer a, an, a human's question and instead make them proactive participants helping to steer your high-value professionals towards the most important areas for their attention. And the way that we're solving that problem is in an experience we're calling the Investor Workbench, where the agent is always

[14:34](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=874s) on tracking what's important to that investor and presenting them with the areas that are most in need of their attention. It looks like an experience like this that relies on the Vault, where a topic is proposed and several pieces of information are presented to the investor to decide whether they should dig deeper or ignore that topic entirely. You can see here the card that an investor sees in order to decide whether to dig in deeper into the same AI impact on currencies question that we looked at earlier in the talk.

[15:05](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=905s) And the investor receives everything they need, an agentic summary if they want a quick scan across the agent's findings. They also get the exact evidence that the agent was drawing on so that they can assess for themselves what was leading to this recommendation. And if they choose to dig deeper, they have a direct connection to the research output from the analytical tools that Aaron was just describing. And when they dig deeper alongside the agent to look at the dynamics driving this question,

[15:35](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=935s) the learnings are automatically captured. The results of their collaboration are automatically captured back in the Vault. And so this for us is the key lesson. The Vault enables compounding because it provides the right evidence, the right ext-- 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. And that makes us go around the circle faster and better.

[16:06](https://www.youtube.com/watch?v=CLve4Nx_RwI&t=966s) And so when we think back to our-- the, the question we began with, what it takes to make an agent work is the right context and the right understanding of what's going on in your world. And so our question for you is: Does your AI know what's going on in the world, and does it remember what your firm has learned about it? Because we are teaching ours to, and if these sounds like an-- sound like an exciting challenge, we'd love to talk. Thank you.
