Descriptions:
Nina Lopatina, head of product at Contextual AI, joins Latent Space to survey the state of context engineering heading into 2026 — covering the evolution from basic RAG pipelines to agentic retrieval, the growing importance of reranking, and why ‘context rot’ has become the dominant failure mode practitioners worry about in production AI systems.
Lopatina explains how Contextual AI’s instruction-following reranker — the first of its kind when released in March — has become increasingly critical in dynamic agentic pipelines, where precision matters more than raw recall at inference time. She also discusses model size adoption trends, citing Open Router data showing medium-sized models (15–70B parameters) gaining share while smaller models plateau, countering some of the hype around on-device AI and Apple Intelligence.
The conversation ranges across benchmark saturation (a research-paper-recreation benchmark was reportedly saturated by Claude Code within days of its release), MCP’s role in the emerging context engineering stack, and which terms like ‘context poisoning’ are failing to stick while ‘context rot’ has become widely adopted shorthand. Lopatina and host Swyx also discuss what a genuine industry benchmark — built on enterprise-scale document corpora rather than toy datasets — would look like. A practical state-of-the-field overview for engineers building RAG or agentic systems in production.
📺 Source: Latent Space · Published December 31, 2025
🏷️ Format: Interview
