Contextual AI CEO Douwe Kiela on RAG 2.0: active retrieval making AI more dynamic and context-aware
Jul 17, 2025 with Douwe Kiela
Key Points
- Contextual AI CEO Douwe Kiela argues that frontier models are now feature-equivalent, making context engineering the true competitive battleground in enterprise AI.
- Production-grade retrieval at enterprise scale requires multi-stage cascading models across millions of documents, rendering long-context windows impractical as a substitute.
- Contextual AI is expanding into code generation and root cause analysis, where retrieval-dependent tasks depend on accessing proprietary documentation outside model weights.
Summary
Douwe Kiela, co-author of the original RAG paper and CEO of Contextual AI, argues that the technology has evolved far beyond its academic origins. The core idea, using retrieval to feed relevant data into generative models, remains unchanged, but enterprise implementation is radically more complex than the research prototype suggested.
The dominant frame in the industry has shifted to what Kiela calls context engineering. Frontier models from Anthropic, OpenAI, and Google are now close enough in raw capability that the differentiating variable is whether the model receives the right context. Getting that context layer right is where Kiela sees the primary competitive opportunity.
The "just use a bigger context window" argument collapses at enterprise scale. Contextual AI's systems operate across millions of documents, a scope that makes long-context models impractical as a substitution for retrieval. Kiela notes that production-grade retrieval is multi-stage, using cascading models of increasing sophistication, not a single lookup step.
Contextual AI currently employs 70 to 80 people and is expanding beyond core document Q&A into adjacent verticals including root cause analysis and code generation. Kiela flags codegen as a particularly retrieval-heavy problem because accurate code synthesis frequently depends on incorporating proprietary technical documentation that sits outside model weights.
Kiela is direct that enterprise search remains unsolved at scale. Data sprawl, inconsistent document formats, and retrieval failures downstream of the model itself are the primary blockers preventing enterprises from extracting value from otherwise capable language models.