Dynamic Context Gathering
Runtime collection of per-instance background under latency, preparability, and comparability constraints — usually the bulk of context-design work.
Once static vs dynamic prompt content has framed the task, the model still lacks subject-specific background (user, document, session). Gathering that material is typically where most context-design and coding time goes: unlike static clarification, it runs while the program is live inside the feedforward pass.
Three constraints shape what is feasible:
- Latency / urgency — non-user or fire-and-forget triggers tolerate slow retrieval; on-demand actions tolerate moderate wait but rarely multi-pass LLM pipelines; in-the-moment completions (typing) make every millisecond of lookup risk the request going stale, so complex non-precomputed retrieval is usually out.
- Preparability — rarely changing facts can be precomputed; under extreme latency pressure, speculatively prepare candidates you might need because there will be no time to fetch them on demand.
- Comparability — gather more than you can pack, then triage. Every item needs a usefulness score (and awareness of dependency / invalidation) so prompt element importance and prompt assembly algorithms can drop the weak ones. Static clarification often earns the highest score because misunderstanding the question dominates; all context is optional in principle — quantify how optional.
Discover candidate sources with dynamic context source discovery, then implement the obvious near sources first and grow outward as the product matures.
Sources
- Prompt Engineering for LLMs (Berryman), ch. 5