Decides what goes into the context window and what is left out, under a hard token limit.
Selects, compresses, orders and truncates the material passed to a model. The unglamorous engineering that determines output quality more than model choice usually does.
Any retrieval-augmented or long-conversation system, which is to say most production deployments.
Short single-turn tasks where everything relevant fits comfortably.
Where material is omitted from context, that omission must be recorded. An answer produced without the relevant evidence is not wrong because the model failed; it is wrong because the evidence was not there, and the trace should say so.
Silent truncation. Content dropped to fit the limit, with the model answering from what remained and no indication anything was missing.
An investigation with forty prior resolutions available and room for eight, where which eight are chosen determines the answer — and the trace records the other thirty-two were excluded.
Two different absences share this shape. Foundational solutions get built whatever the domain, so no domain links them; the rest are solutions this domain genuinely does not reach for. v_ai_solutions_unlinked separates the two.