Gemini 1.5 made long context the main event

Gemini 1.5 Pro made long context feel like a product primitive instead of a footnote.

That does not mean every request should include every document. Retrieval still decides what is relevant and permitted. The difference is that useful source material does not have to be compressed immediately into a summary that may discard the one clause that matters.

For a policy comparison, the workflow can stay explicit:

request: compare current and old refund language
retrieval: select the full policy sections
context: keep both sections intact
answer: cite the clause that differs

The input manifest is part of the evidence:

{
  "request_id": "refund-policy-184",
  "sources": [
    {
      "path": "docs/billing/current-refund-policy.md",
      "role": "current_source_of_truth",
      "selected_unit": "full_section"
    },
    {
      "path": "docs/archive/refund-policy-2022.md",
      "role": "historical_comparison",
      "selected_unit": "full_section"
    }
  ],
  "question": "Which clause controls annual-plan refunds?"
}

That record shows whether the model saw the right sources and whether the comparison was fair. It also preserves source shape: a policy section, contract clause, and changelog entry are not interchangeable.

A short context window can omit the controlling clause before reasoning begins. A large one can keep the evidence available even when the answer is wrong, which makes the failure easier to inspect.

That was the practical shift. Long context stopped sounding like an experiment and started looking like a system component.