1
00:00:01,441 --> 00:00:03,181
Just waiting for our slides.

2
00:00:03,641 --> 00:00:04,581
There we go.

3
00:00:05,141 --> 00:00:06,041
Good morning, everyone.

4
00:00:06,521 --> 00:00:12,361
I'm Sam Green, team lead for Applied AI at Bridgewater Associates, and this is my colleague,

5
00:00:12,521 --> 00:00:13,061
Aaron Linsky.

6
00:00:13,061 --> 00:00:18,781
and today we're here to talk to you about the vault, the data platform that we've built at

7
00:00:18,781 --> 00:00:23,481
Bridgewater to turn unstructured data into a first-class asset in our investment process.

8
00:00:24,361 --> 00:00:28,681
And as part of the talk this morning, we'd like to share three lessons that we've learned building

9
00:00:28,681 --> 00:00:35,401
the vault. The first is that humans and agents have very different needs from a data platform,

10
00:00:36,071 --> 00:00:41,141
and so you have to build for those different needs separately. The second is that the greatest

11
00:00:41,141 --> 00:00:45,781
impact occurs when a single tool can meet both of these sets of needs at the same time.

12
00:00:46,241 --> 00:00:51,961
And the third is that the real prize is when you can build the foundations needed to make

13
00:00:51,961 --> 00:00:57,381
agents a proactive participant in your investment process or whatever you're doing with your

14
00:00:57,381 --> 00:01:02,981
tools instead of a reactive tool for your humans to use.

15
00:01:04,801 --> 00:01:08,761
Before I dive deeper into talking about the vault, I wanted to start with some background

16
00:01:08,761 --> 00:01:09,421
about Bridgewater.

17
00:01:10,221 --> 00:01:15,541
Over the last 50 years, Bridgewater has been in the constant pursuit of the world's best

18
00:01:15,541 --> 00:01:21,081
understanding of how markets and economies work, and turning that into a systematic encoding

19
00:01:21,081 --> 00:01:24,101
that we can use to stress test our logic over time.

20
00:01:24,941 --> 00:01:31,161
And what originally began as a set of if-then rules on a yellow legal pad evolved into a

21
00:01:31,161 --> 00:01:35,741
continuous compounding process with humans at the center constantly looking for the next

22
00:01:35,741 --> 00:01:37,041
thing that's changing in markets.

23
00:01:37,041 --> 00:01:42,061
And more recently, we've been working on building AYA, the artificial investment associate,

24
00:01:42,401 --> 00:01:45,561
which puts a machine at the center of this continuous process.

25
00:01:46,261 --> 00:01:50,901
Across all these phases of evolution, we've had a constant philosophy in mind, which is

26
00:01:50,901 --> 00:01:55,201
that we want to be constantly looking at how the world is changing and what's going on

27
00:01:55,201 --> 00:02:02,241
in the world, noticing what dynamics are not yet encoded in our understanding, and use that

28
00:02:03,181 --> 00:02:05,701
learning to improve and compound for the future.

29
00:02:07,561 --> 00:02:09,281
For most of Bridgewater's history,

30
00:02:09,521 --> 00:02:11,321
that meant building and overseeing

31
00:02:11,321 --> 00:02:13,601
time series models on structured data.

32
00:02:14,281 --> 00:02:17,521
And that structured data needed to be carefully built

33
00:02:17,521 --> 00:02:20,541
so that it could be safe for use in our research process

34
00:02:20,541 --> 00:02:23,981
and aligned to our internal ontology and world model

35
00:02:23,981 --> 00:02:25,441
so that we could get great leverage

36
00:02:25,441 --> 00:02:27,281
out of using the data that we had built.

37
00:02:28,421 --> 00:02:30,601
With the emergence of language models,

38
00:02:30,641 --> 00:02:31,981
we had a huge new opportunity

39
00:02:31,981 --> 00:02:34,941
to engage with a much wider world of data,

40
00:02:36,161 --> 00:02:39,001
news and research, transcripts, policy documents,

41
00:02:39,141 --> 00:02:41,401
and all of the internal conversation

42
00:02:41,401 --> 00:02:44,061
that was previously just in the ether.

43
00:02:44,241 --> 00:02:46,861
So the internal memos and meeting transcripts

44
00:02:46,861 --> 00:02:48,341
and debate happening inside the firm.

45
00:02:48,881 --> 00:02:51,341
We quickly realized that all of that data

46
00:02:51,341 --> 00:02:54,681
needed the exact same treatment to be fit for purpose.

47
00:02:54,961 --> 00:02:57,201
It needed to be carefully cleaned, governed,

48
00:02:57,201 --> 00:03:01,861
and made ready for use in both point-in-time research questions

49
00:03:01,861 --> 00:03:09,081
today and also our systematic study over time. And when we looked out at the landscape of data

50
00:03:09,081 --> 00:03:14,021
vendors and data platform providers, there was nothing that provided us with all of the control

51
00:03:14,021 --> 00:03:22,101
that we needed to build and maintain our data advantage. And so we built the vault, which is our

52
00:03:22,101 --> 00:03:29,721
repository of external and internal unstructured data, and the platform that indexes, organizes,

53
00:03:29,721 --> 00:03:33,481
and makes all of it ready for use by our human investors and agents.

54
00:03:34,101 --> 00:03:36,941
Today, the vault has more than 100 million documents in it,

55
00:03:37,281 --> 00:03:40,301
with thousands more added every day, minute by minute,

56
00:03:40,581 --> 00:03:44,641
from 135 external sources, and again, growing all the time,

57
00:03:44,841 --> 00:03:47,861
as well as all of the data from inside of Bridgewater.

58
00:03:48,161 --> 00:03:52,221
And crucially, all that data is not just dumped on our investors' desks.

59
00:03:52,441 --> 00:03:56,641
It is carefully aligned to our internal ontology

60
00:03:56,641 --> 00:04:01,281
with enrichments like the countries, concepts, markets,

61
00:04:01,801 --> 00:04:05,301
and most importantly, time period discussed in the document.

62
00:04:06,601 --> 00:04:09,781
To bring this to life, I want to start with a quick motivating example.

63
00:04:10,121 --> 00:04:16,441
So what I did here is I asked a fairly basic representative investment question

64
00:04:16,441 --> 00:04:18,261
to an enterprise chat client.

65
00:04:18,581 --> 00:04:22,761
I asked, how is AI infrastructure investment impacting the Korean won,

66
00:04:22,961 --> 00:04:25,221
and do market participants expect this to continue?

67
00:04:25,221 --> 00:04:32,361
you. As you are all familiar with, the agent goes to the web and looks for resources it

68
00:04:32,361 --> 00:04:38,361
can draw on because I have explicitly asked it to go find other people's perspectives

69
00:04:38,361 --> 00:04:45,181
on the question. You can see it zooms in and finds 15 relevant looking resources to draw

70
00:04:45,181 --> 00:04:51,801
on before it gives me back its answer. When I zoomed in on the answer it gave back, a

71
00:04:51,801 --> 00:04:52,821
A couple of things stood out to me.

72
00:04:53,081 --> 00:04:58,441
The first is that even though it found 15 relevant-looking sources, it only picked one

73
00:04:58,441 --> 00:05:00,621
or two to answer my question.

74
00:05:01,121 --> 00:05:06,141
And the second is down at the end of this response, where the agent itself acknowledges

75
00:05:06,141 --> 00:05:07,581
that it hasn't answered my question.

76
00:05:08,281 --> 00:05:14,061
It didn't find enough documents on the open web to establish a market consensus and meet

77
00:05:14,061 --> 00:05:15,821
the parameters that I laid out.

78
00:05:17,001 --> 00:05:19,761
When I double-clicked even further, the problem got even worse.

79
00:05:21,041 --> 00:05:33,601
One of the notes that it primarily drew on to answer my question was published way back in July of 2026, even though I asked this question and captured this answer just at the end of last week in early October.

80
00:05:34,681 --> 00:05:40,101
In the age of AI and in investing, three months is a massive amount of time.

81
00:05:40,481 --> 00:05:45,201
It's not even close to good enough to draw on this resource to answer my question.

82
00:05:46,721 --> 00:05:52,141
When I ask this exact same query attached to the vault, something totally different happens.

83
00:05:52,581 --> 00:05:57,041
The first is that the connection to the vault helps to prompt the agent that it should look

84
00:05:57,041 --> 00:05:59,401
for a specific period of time when answering my question.

85
00:05:59,581 --> 00:06:01,921
So it zooms in on October 1st to October 8th.

86
00:06:01,921 --> 00:06:08,241
It finds a much wider array of resources, 52 candidate documents that are relevant to

87
00:06:08,241 --> 00:06:13,381
the query over this specific period of time because of the vault's much more significant

88
00:06:13,381 --> 00:06:16,181
corpus of content that's relevant to investing questions.

89
00:06:17,661 --> 00:06:22,341
when it ultimately answers my question, it draws on seven instead of only one or two,

90
00:06:23,921 --> 00:06:28,861
gives me back a rich answer that the agent is more confident is going to address my question.

91
00:06:30,041 --> 00:06:35,621
And so for us, this is the big lesson from building the vault to motivate the talk today,

92
00:06:35,801 --> 00:06:39,361
which is that it takes great data and the right connectors to make agents impactful.

93
00:06:40,421 --> 00:06:43,941
And I'm going to hand it over to Aaron now to talk about some of the product engineering

94
00:06:43,941 --> 00:06:45,061
takeaways from building the vault.

95
00:06:47,261 --> 00:06:47,601
Thanks, Sam.

96
00:06:48,621 --> 00:06:49,621
So hi, I'm Aaron.

97
00:06:49,801 --> 00:06:51,321
I run product engineering at Bridgewater.

98
00:06:51,561 --> 00:06:54,781
I'm going to take us through a few of our learnings from building the vault over the

99
00:06:54,781 --> 00:06:55,541
past few years.

100
00:06:55,541 --> 00:07:00,441
I'm going to start with one that might seem obvious, but I still think it's worth saying,

101
00:07:00,661 --> 00:07:02,181
which is search is hard.

102
00:07:04,361 --> 00:07:08,081
But it gets even harder if we're designing for two different consumers.

103
00:07:08,341 --> 00:07:10,961
In this case, we're designing for humans and agents.

104
00:07:11,621 --> 00:07:13,881
I'm going to go over the difference here, starting with us.

105
00:07:14,481 --> 00:07:19,701
So humans need precision and a high signal-to-noise ratio because we can only read so much.

106
00:07:20,001 --> 00:07:25,181
Even a fast reader manages about 300 words a minute, and personally, I'm much slower.

107
00:07:26,481 --> 00:07:28,761
Agents are a different kind of reader.

108
00:07:29,141 --> 00:07:30,901
They're far less sensitive to precision.

109
00:07:31,321 --> 00:07:33,201
If the data is there, they'll find it.

110
00:07:33,981 --> 00:07:37,261
And basically anything you hand to them will get attended to.

111
00:07:37,601 --> 00:07:45,641
They scale super quickly and they consume about 100 times faster, 30,000 words per minute.

112
00:07:46,021 --> 00:07:49,061
And so we need to build Vault for both of these.

113
00:07:49,601 --> 00:07:54,501
And when we're thinking about the dimensions that matter, there are these five on the board.

114
00:07:54,501 --> 00:08:01,541
recall, precision, filtering, entitlements, and scale. Some of them matter to people,

115
00:08:01,801 --> 00:08:06,501
some matter more to agents, many to both. I'm going to go deep on three of these,

116
00:08:06,841 --> 00:08:13,261
precision for people, scale for agents, and then filtering for both. So, precision.

117
00:08:14,501 --> 00:08:20,461
Once we've filled the vault, which was a heroic effort by our data and legal teams,

118
00:08:21,081 --> 00:08:23,421
Recall was no longer a pressing issue.

119
00:08:23,621 --> 00:08:27,101
However, the number one complaint was vault search is bad.

120
00:08:27,941 --> 00:08:30,361
As a product owner, it's painful to hear that,

121
00:08:30,561 --> 00:08:32,701
but my response was, let's benchmark it.

122
00:08:33,041 --> 00:08:34,161
And that's what we did.

123
00:08:34,841 --> 00:08:38,401
We collected questions and found the best documents by hand,

124
00:08:38,641 --> 00:08:40,761
including ones that looked helpful but weren't.

125
00:08:41,441 --> 00:08:43,861
In total, it was about 200 examples,

126
00:08:44,001 --> 00:08:45,261
and it took us less than a week.

127
00:08:45,401 --> 00:08:48,201
It was not fun or very glamorous,

128
00:08:48,441 --> 00:08:49,361
but it was very helpful.

129
00:08:50,461 --> 00:08:52,621
And after we ran it, it was very clear.

130
00:08:52,761 --> 00:08:53,821
Search was bad.

131
00:08:54,561 --> 00:08:57,621
The key metric we were looking at is precision at 10,

132
00:08:57,861 --> 00:09:01,421
which is of the top 10 results, how many are the right documents.

133
00:09:02,601 --> 00:09:05,141
And we were hitting about 60%.

134
00:09:05,141 --> 00:09:07,121
So the people were right.

135
00:09:07,241 --> 00:09:07,741
It was bad.

136
00:09:08,221 --> 00:09:11,421
But now we had a way to measure all of the improvements.

137
00:09:11,741 --> 00:09:14,061
And here on the board are five of them.

138
00:09:14,181 --> 00:09:16,561
And by the time we got done with agentic search,

139
00:09:16,561 --> 00:09:21,941
we were up in the 90-95% recall at 10, or precision at 10.

140
00:09:23,701 --> 00:09:25,141
All right, so that was precision.

141
00:09:25,801 --> 00:09:26,841
Now on to filtering.

142
00:09:27,901 --> 00:09:31,261
Like anyone, we want the most relevant documents.

143
00:09:31,441 --> 00:09:33,801
But when we looked at the production traces,

144
00:09:34,101 --> 00:09:37,081
what we were seeing is that almost always there was some condition,

145
00:09:37,361 --> 00:09:41,301
either as of a certain date or from a certain source

146
00:09:41,301 --> 00:09:42,741
or about a specific asset.

147
00:09:43,521 --> 00:09:46,481
And all of that requires advanced filtering.

148
00:09:46,561 --> 00:09:49,941
And that's where we ran into our first major performance issue.

149
00:09:51,161 --> 00:09:55,201
So I'm going to quickly go through this, a little bit technical here.

150
00:09:55,441 --> 00:09:59,541
There's classically two ways to filter a vector database.

151
00:10:00,061 --> 00:10:02,621
Post-filtering, we'll use that vector index first,

152
00:10:02,661 --> 00:10:04,921
and then filter the top results by metadata.

153
00:10:05,401 --> 00:10:08,441
It's fast, but recall could be pretty low,

154
00:10:08,481 --> 00:10:11,981
because the globally relevant results might not match your filters.

155
00:10:13,061 --> 00:10:15,421
On the other side, we have pre-filtering,

156
00:10:15,421 --> 00:10:19,921
where we'll filter first and then calculate distances for every one of those matches.

157
00:10:20,341 --> 00:10:23,961
The advantage there is that the recall is very good, but it can be very slow,

158
00:10:24,061 --> 00:10:25,241
and we needed something in between.

159
00:10:26,741 --> 00:10:30,241
So that's what led us to try and then ultimately switch to turbopuffer.

160
00:10:30,421 --> 00:10:32,781
Their approach, which they call native filtering,

161
00:10:33,001 --> 00:10:36,881
looks only in the nearest clusters that contain matches and then scores just those.

162
00:10:37,141 --> 00:10:40,921
What you end up getting is something that's fast, like post-filtering,

163
00:10:41,081 --> 00:10:44,161
but with close to the same recall as pre-filtering.

164
00:10:46,141 --> 00:10:47,941
Lastly, I'm going to talk about scale.

165
00:10:50,061 --> 00:10:51,721
scale for us is pretty pedestrian.

166
00:10:51,981 --> 00:10:56,561
100 queries an hour, we have 300 users, they're going to search every so often.

167
00:10:57,261 --> 00:11:01,361
It takes our users minutes to read the docs that they're receiving.

168
00:11:01,961 --> 00:11:05,721
Agents on the other hand are a few orders of magnitude beyond that.

169
00:11:06,221 --> 00:11:10,641
Easily 100 plus queries a second and agent swarms can hit thousands easily.

170
00:11:11,521 --> 00:11:14,801
And they obviously will read, as I said earlier, will read docs in seconds.

171
00:11:14,801 --> 00:11:16,981
So we needed something that could scale as well.

172
00:11:17,641 --> 00:11:20,201
And again, turbopuffer handled that easily too.

173
00:11:21,581 --> 00:11:23,801
So we were covering all five dimensions.

174
00:11:24,001 --> 00:11:26,721
I didn't talk about entitlements, but I'll be talking about that later today.

175
00:11:27,341 --> 00:11:28,541
But we covered it all.

176
00:11:28,861 --> 00:11:30,061
It unlocked new workflows.

177
00:11:31,161 --> 00:11:34,361
But it didn't change how our researchers worked.

178
00:11:34,561 --> 00:11:36,601
They still weren't coming to Vault every morning.

179
00:11:37,001 --> 00:11:39,721
And that's my second lesson that I want to share here,

180
00:11:39,841 --> 00:11:42,381
is that no matter how amazing the tools are,

181
00:11:42,461 --> 00:11:44,581
we can't just drop them on our investors' desks.

182
00:11:44,801 --> 00:11:50,841
and expect them to be used. We need to migrate their old habits into a new form.

183
00:11:51,441 --> 00:11:56,961
And so we looked at what they were already doing. Many of them were living in email.

184
00:11:57,221 --> 00:12:02,001
The research came straight from the brokers. They had subscribed lists. They had rules

185
00:12:02,001 --> 00:12:07,281
that filtered everything into folders. It worked. It was familiar. And they knew that

186
00:12:07,281 --> 00:12:11,601
they weren't missing anything. So we had to build all of these features into the vault.

187
00:12:12,421 --> 00:12:13,581
That's exactly what we did.

188
00:12:13,801 --> 00:12:17,341
Everything searchable, the subscribed lists became saved searches,

189
00:12:17,821 --> 00:12:20,301
and folders became saved collections.

190
00:12:20,581 --> 00:12:22,561
It's the same habits, just a new home.

191
00:12:22,961 --> 00:12:25,621
And once we had them there, we can open up new habits.

192
00:12:25,961 --> 00:12:29,801
So they can ask questions across all the data, as Sam showed.

193
00:12:30,081 --> 00:12:32,421
They can follow dynamic feeds by topic,

194
00:12:32,521 --> 00:12:36,281
and they can hand those collections over to their agents to take further action.

195
00:12:36,501 --> 00:12:38,901
And that's when we clearly had product market fit.

196
00:12:40,641 --> 00:12:45,521
And once we had that fit, the users wanted it everywhere, so we put it in chat, we made

197
00:12:45,521 --> 00:12:49,661
a phone app, it's in Slack in their IDE, and of course we availed all of that to the agents

198
00:12:49,661 --> 00:12:54,261
as well, including a CLI, because agents love CLIs.

199
00:12:55,561 --> 00:12:58,901
And so the main question here is, okay, so what does this enable?

200
00:12:59,861 --> 00:13:04,161
Sam's going to talk about what we're building right now, but first and foremost, it's helping

201
00:13:04,161 --> 00:13:06,401
us to move around that research circle faster.

202
00:13:07,461 --> 00:13:18,881
With our pocket analyst, which is our first AI analyst, it can do hours of research in just a few minutes for a few dollars, and that's relying on the Vault.

203
00:13:19,441 --> 00:13:23,081
And then we're expanding that now with the researcher, which is the next generation.

204
00:13:23,081 --> 00:13:26,121
It's the same idea stretched across days or weeks.

205
00:13:27,541 --> 00:13:32,421
We're also able to train custom models to more intelligently reason over the Vault's data,

206
00:13:32,621 --> 00:13:37,801
and use those together with other foundation models to feed systematic decision-making processes.

207
00:13:38,201 --> 00:13:43,781
The Forecaster is our first one, which produces AI-generated macro-relevant predictions.

208
00:13:44,921 --> 00:13:48,481
then we also have built investment signals from the Vault,

209
00:13:48,741 --> 00:13:52,361
turning unstructured data into structured, tradable strategies.

210
00:13:52,981 --> 00:13:56,761
Now I will hand it to Sam to take us through Vault's latest capability.

211
00:14:00,841 --> 00:14:05,841
like I alluded to at the beginning of the talk, the real prize is when you can take

212
00:14:05,841 --> 00:14:11,941
agents from a reactive resource that can be ready to answer a human's question and instead

213
00:14:11,941 --> 00:14:17,061
make them proactive participants helping to steer your high value professionals towards

214
00:14:17,061 --> 00:14:19,161
the most important areas for their attention.

215
00:14:19,921 --> 00:14:24,401
The way we're solving that problem is in an experience we're calling the investor workbench,

216
00:14:24,761 --> 00:14:31,401
where the agent is always on, tracking what's important to that investor and presenting

217
00:14:31,401 --> 00:14:34,401
them with the areas that are most in need of their attention.

218
00:14:35,681 --> 00:14:40,441
looks like an experience like this that relies on the vault, where a topic is proposed

219
00:14:40,441 --> 00:14:45,441
and several pieces of information are presented to the investor to decide whether they should

220
00:14:45,441 --> 00:14:47,641
dig deeper or ignore that topic entirely.

221
00:14:48,281 --> 00:14:57,661
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.

222
00:14:58,181 --> 00:15:01,101
And the investor receives everything they need.

223
00:15:01,321 --> 00:15:13,141
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 they can assess for themselves what was leading to this recommendation.

224
00:15:14,101 --> 00:15:18,301
And if they choose to dig deeper, they have a direct connection to the research output

225
00:15:18,301 --> 00:15:20,021
from the analytical tools that

226
00:15:20,021 --> 00:15:21,021
Aaron was just describing.

227
00:15:21,721 --> 00:15:35,501
And when they dig deeper alongside the agent to look at the dynamics driving this question, the learnings are automatically captured, the results of their collaboration are automatically captured back in the vault.

228
00:15:37,061 --> 00:15:38,761
And so this for us is the key lesson.

229
00:15:39,101 --> 00:15:45,081
The vault enables compounding because it provides the right evidence, the right internal and

230
00:15:45,081 --> 00:15:49,881
external information at the fingertips of our investors and agents, and it automatically

231
00:15:49,881 --> 00:15:54,141
captures the synthesis that they generate and puts it back in the corpus for the next

232
00:15:54,141 --> 00:15:54,581
turn.

233
00:15:55,501 --> 00:15:57,921
And that makes us go around the circle faster and better.

234
00:15:59,261 --> 00:16:04,521
And so when we think back to the question we began with, what it takes to make an agent

235
00:16:04,521 --> 00:16:08,301
work is the right context and understanding of what's going on in your world.

236
00:16:09,161 --> 00:16:13,521
Our question for you is, does your AI know what's going on in the world and does it remember

237
00:16:13,521 --> 00:16:14,841
what your firm has learned about it?

238
00:16:15,361 --> 00:16:17,421
Because we are teaching ours too.

239
00:16:17,781 --> 00:16:20,561
And if these sound like an exciting challenge, we'd love to talk.

240
00:16:20,901 --> 00:16:21,421
Thank you.