{
  "video": {
    "id": "CLve4Nx_RwI",
    "title": "How Bridgewater Turned 100 Million Documents Into an Edge - Sam Green & Aaron Linsky, Bridgewater",
    "duration": 1005,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "CLve4Nx_RwI",
    "title": "How Bridgewater Turned 100 Million Documents Into an Edge - Sam Green & Aaron Linsky, Bridgewater",
    "one_liner": "Bridgewater's Vault holds 100M+ unstructured documents aligned to the firm's internal ontology, including the time period each document discusses. Their lesson is that agents need great data and the right connectors to be useful: the same Korean won question pulled 52 candidate documents and cited 7 through the Vault, against 15 found and 1-2 used on the open web.",
    "summary": "Sam Green and Aaron Linsky describe the Vault, Bridgewater's own platform for external and internal unstructured data. They argue three lessons. First, humans and agents need different things from a data platform, so you build for each separately. Second, the biggest impact comes when one tool serves both at once. Third, the real prize is agents that take part in the investment process proactively instead of waiting to be asked. They show this with search benchmarking, filtered vector search on Turbopuffer, rebuilding investors' email habits inside the Vault, and an always-on 'investor workbench' that writes its findings back into the corpus.",
    "key_points": [
      "The Vault holds more than 100 million documents from 135 external sources plus Bridgewater's internal memos, meeting transcripts and debate, with thousands added every day. Each document is enriched against the internal ontology with countries, concepts, markets and, most importantly, the time period it discusses.",
      "Demo question: 'How is AI infrastructure investment impacting the Korean won?' An enterprise chat client on the open web found 15 sources, used only 1-2, admitted it could not establish a market consensus, and leaned mainly on a note from July 2026 when the question was asked in early October. Connected to the Vault, the agent narrowed to October 1-8, found 52 candidate documents and drew on 7.",
      "Humans read about 300 words a minute and need precision and a high signal-to-noise ratio. Agents read about 30,000 words a minute (100x faster), care far less about precision, and attend to anything they are given. The five dimensions they designed for were recall, precision, filtering, entitlements and scale.",
      "Users said 'Vault search is bad', so the team built a benchmark: about 200 questions with the best documents picked by hand, including ones that looked helpful but weren't. It took less than a week. Precision@10 started around 60% and reached 90-95% after five improvements, the last being agentic search.",
      "Production traces showed almost every query carried a condition: as of a date, from a source, or about an asset. Post-filtering a vector index is fast but loses recall; pre-filtering keeps recall but is slow. They switched to Turbopuffer, whose 'native filtering' searches only the nearest clusters that contain matches, giving close to pre-filter recall at close to post-filter speed.",
      "Human load is about 100 queries an hour from 300 users. Agents easily reach 100+ queries a second, and agent swarms reach thousands. Turbopuffer handled that scale too.",
      "Better search alone didn't change how researchers worked. Adoption came from moving their email habits into the Vault: broker subscription lists became saved searches and folders became saved collections. After that came dynamic topic feeds and handing collections to agents, which is when they had product-market fit. The Vault then went into chat, a phone app, Slack, the IDE, and a CLI for agents.",
      "Products built on the Vault: Pocket Analyst does hours of research in minutes for a few dollars; Researcher is the same idea stretched across days or weeks; Forecaster is a custom model that makes macro-relevant predictions; and investment signals turn unstructured data into structured, tradable strategies. The Investor Workbench is an always-on agent that proposes topics as cards with an agentic summary and the exact evidence, and automatically writes the results of investor-agent collaboration back into the Vault."
    ],
    "takeaways": [
      "Benchmark search before trying to fix it. About 200 hand-labelled queries, including tempting false positives, take under a week and turn 'search is bad' into a precision@10 number you can improve against.",
      "Enrich documents with metadata, especially the time period discussed, and choose a vector store that filters without giving up recall or speed. Real queries are nearly always conditioned on date, source or asset.",
      "Design for two consumers on purpose: precision and low noise for humans, throughput (100s to 1000s of queries a second) and recall for agents, plus agent-friendly interfaces like a CLI.",
      "To drive adoption, rebuild users' existing workflows inside the new tool (subscriptions become saved searches, folders become collections) before offering new habits.",
      "Close the loop: write what humans and agents produce together back into the corpus so the system compounds, and move agents from answering questions to proactively surfacing what needs attention, with the evidence attached."
    ],
    "topics": [
      "unstructured data",
      "search",
      "vector databases",
      "agents",
      "evals",
      "rag",
      "finance",
      "product adoption"
    ],
    "tools": [
      "Bridgewater Associates",
      "the Vault",
      "AiA (Artificial Investment Associate)",
      "Turbopuffer",
      "Slack",
      "Pocket Analyst",
      "Researcher",
      "Forecaster",
      "Investor Workbench"
    ],
    "quotes": [
      {
        "text": "it takes great data and the right connectors to make agents impactful.",
        "at": "06:42",
        "url": "https://www.youtube.com/watch?v=CLve4Nx_RwI&t=402s"
      },
      {
        "text": "It's the same habits, just a new home.",
        "at": "12:27",
        "url": "https://www.youtube.com/watch?v=CLve4Nx_RwI&t=747s"
      },
      {
        "text": "including a a CLI because agents love CLIs.",
        "at": "12:58",
        "url": "https://www.youtube.com/watch?v=CLve4Nx_RwI&t=778s"
      },
      {
        "text": "does your AI know what's going on in the world? And does it remember what your firm has learned about it?",
        "at": "16:16",
        "url": "https://www.youtube.com/watch?v=CLve4Nx_RwI&t=976s"
      }
    ],
    "words": 3266
  },
  "summary_url": "/#CLve4Nx_RwI",
  "transcript": {
    "html": "/transcripts/CLve4Nx_RwI.html",
    "txt": "/transcripts/CLve4Nx_RwI.txt",
    "vtt": "/transcripts/CLve4Nx_RwI.vtt"
  }
}