Bridgewater's data platform: humans need precision, agents need scale.

The motivating example. Sam Green opened with a simple query about AI's impact on the Korean won. A generic agent searched the web, found 15 sources, but only used two, one of which was three months old—an eternity in markets. The same query against Bridgewater's Vault platform returned 52 relevant documents from the correct time window and cited seven. The lesson: great data and proper connectors make agents impactful.

The platform. Bridgewater built Vault to house over 100 million unstructured documents from 135 external sources plus internal memos and transcripts. Every document is enriched with internal ontology tags—countries, concepts, markets, and time periods—so both humans and agents can filter and retrieve with high relevance. This curation turns raw data into a first-class asset for investment research.

Designing for two consumers. Aaron Linsky described the fundamental challenge: humans and agents read at vastly different speeds (300 vs 30,000 words per minute) and have different tolerances. Humans need high precision; agents can consume large result sets. Bridgewater benchmarked precision at 10—top 10 results correct—and found it at 60%. After iterative improvements with agentic search, they reached 90-95%.

The filtering bottleneck. Most vector database filtering falls into post-filtering (fast but low recall) or pre-filtering (accurate but slow). Bridgewater needed a middle ground for time-sensitive queries. They switched to Turbopuffer, which uses native filtering—scoring only nearest clusters that match metadata conditions. The result: speed near post-filtering with recall close to pre-filtering.

Habit migration, not just tooling. The team learned a great platform won't get used if it breaks existing workflows. Researchers lived in email with subscribed lists and folders. Vault replicated these as saved searches and collections. Only after matching old habits could Bridgewater introduce new capabilities like agent-driven question answering and dynamic feeds. This was the real product-market fit moment.

Proactive agents and compounding. The latest capability, the investor workbench, moves agents from reactive to proactive—surfacing topics needing attention. The agent proposes topics with summaries and evidence; the investor can dig deeper alongside it. Every interaction gets captured back into Vault, creating a compounding loop where the platform improves with each use.