Bridgewater's 100M permissioned docs: turbopuffer's native filtering hits 120ms P90 on 400M vectors
AI Engineer · youtube · 2026-10-25
AI Engineer talk with Aaron Linsky (head of product engineering, Bridgewater Associates) and Nikhil Benesch (CTO, turbopuffer) on permission-aware vector search for AI analysts:
- Problem: Bridgewater's AI analyst needs internal research docs, but every email, memo and note carries its own permissions. Their Vault holds 100M permissioned documents, which broke the first search setup
- Trade-offs: post-filtering loses recall; pre-filtering is too slow; brute force mathematically fails at 400M vectors
- Fix: turbopuffer's "native filtering" runs permission filters natively over a clustered vector index, getting both recall and speed
- Result: 120ms P90 search latency across 400M vectors / 100M documents
Chapters cover how vector indexes work, why permissions break them, filter strategy comparisons, and writing permission filters in practice — valuable for engineers building enterprise RAG.
More from Infra
- Free iPhone idea: 42-month 0% Apple Lease paid off by renting idle compute to AI — SuB8u · 2026-10-25
- 4B model fine-tuned 27 min on one A100 lifts letter-exam accuracy to 74.5% — MaziyarPanahi · 2026-10-25
- Starlink in-flight WiFi hits 408Mbps, ~30% faster than median US home broadband — elonmusk · 2026-10-25
- Musk: Mine lunar silicon to build AI satellites — Earth launches can't scale to 500-1,000 TW a year — elonmusk · 2026-10-25
- Router dev crowdsources production model picks beyond the big four as small-model gap narrows — Remarkable-Guille · 2026-10-25
- Ultrapolis dev breaks down what it takes to render every frame at 60fps — mflux · 2026-10-25