Why we're looking at this
Every few months a larger context window ships, and every time, someone asks whether retrieval-augmented generation is now obsolete. We keep both in production, so we care about the actual tradeoffs rather than the headline.
What we're seeing
- Long context is genuinely better for tasks that need to reason across an entire document coherently — contracts, long transcripts, full codebases.
- Retrieval remains cheaper and faster at scale, and it's the only practical option when the underlying knowledge base is larger than any context window and changes frequently.
- Retrieval gives you a controllable audit trail — you can point to exactly which chunk supported an answer. That property alone made it the right choice for Physiolaxy's evidence-based protocol engine, where every recommendation needs a traceable source.
Open questions we're still chasing
Hybrid approaches — retrieval to narrow the field, then a long-context pass over the retrieved set — are where we're spending most of our attention right now, since they seem to combine the audit trail of retrieval with the coherence of long context.