How to Connect AI to Billions of Legal Documents
Official Schedule Context
- Date/time: 2026-06-29 · 2:25pm-2:45pm
- Track/room: Search & Retrieval · Track 3
- Speaker(s): Simon Eskildsen, Jacob Lauritzen
- Session type/status: session · confirmed
Official Description
Legora’s foundational engineering challenge is connecting frontier LLMs to billions of legal
documents so the models can efficiently solve end-to-end legal workflows without burning extra
tokens. We’ll share the retrieval architecture we built with turbopuffer that achieves: 1. Strict
data isolation across millions of legal cases in a very security-conscious domain 2. Predictable
search performance (<100ms p90 latency) on large contexts 3. High retrieval quality (95%+ recall@10)
with fewer agent loops We’ll retrospect on two architectures that failed to achieve all 3 (and why),
and the key design factors that make the current solution work at our scale. Practical takeaways
include: - How to evaluate per-tenant vs shared-index retrieval under strict data isolation - How to
efficiently index and retrieve context to maximize relevance per input token - How to build a highly
intelligent AI application when your inference budget is constrained
Related YouTube Video
Agents need more than a chat - Jacob Lauritzen, CTO Legora (speaker-match related prior/adjacent AI Engineer video; captions: English auto-captions).
Transcript Status
Related video transcript availability: English auto-captions. Treat this as supporting context, not a recording of this exact scheduled session unless later confirmed. Not fetched yet.
People
Notes
- Pending transcript synthesis when an official recording or confirmed matching video is available.
Supporting Slides
- youtube XNtkiQJ49Ps slides — extracted from the related public AI Engineer video.
Slide Evidence
- Slide-only cropped deck: youtube XNtkiQJ49Ps dense slides (7 viable slide images).
- Related slide/OCR pages:
- youtube XNtkiQJ49Ps dense slides
- youtube XNtkiQJ49Ps reconstructed slides
- youtube XNtkiQJ49Ps slides
- Slide-derived terms:
legora,chat,than,company,jacob,lauritzen,trust,searching,reading,braintrust,workos,openal,files,file,humans,alengineer,vecoea,collard