Context engineering today often means retrieving the right chunk, but the question itself also needs careful treatment. In a typical insurance example, the raw string “What is the maximum coverage amount? Don’t confuse it with the deductible, they’re often listed together.” is split into four signals—topic, negative cue, expected shape, structural hint—each destined for a different downstream step.

These signals become a typed row written by the question‑parsing brick of Enterprise Document Intelligence. The parser uses a fixed PARSE_QUESTION_SYSTEM_PROMPT, a compact doc‑context JSON, a few‑shot block, and the raw user question to build a deterministic row of five columns that no downstream call must re‑parse.

LangChain’s context‑engineering taxonomy maps four canonical strategies—write, select, compress, isolate—to the four typed pieces the parser emits. The write strategy produces the Parsed Question contract; compress creates a smaller retrieval brief; select generates a deterministic generation brief; isolate handles clarification requests.

Each downstream brick consumes only the fields it needs: retrieval uses the brief to anchor searches, generation reads the brief to choose the model tier and chunk strategy, and the dispatcher logs decisions. This disciplined, typed approach prevents mis‑routing and makes the pipeline reproducible and auditable.