Dynamic Context Source Discovery
Find what runtime context to gather by mind-mapping the question and ranking sources by application proximity and stability.
Dynamic context gathering needs a systematic way to invent candidate facts before wiring retrieval. Two complementary moves:
Mind-map the target question. Put the ask in the center and vary words or aspects ("What book shall I read next?" → drop "next", vary "I") until follow-up questions appear ("What did I read last?", "Did I like it?"). This surfaces context you would want if you had it; obtaining some items may be infeasible or deferred to a later version.
Inventory what you can actually gather, then assess relevance. Sort candidates on either axis (pick one; both are optional):
- Proximity to the application (nearer ≈ easier): on-screen / system state → already-saved profiles → activity the app could start recording → public APIs → permissioned systems or direct user questions. Farther sources must earn more usefulness to justify cost.
- Stability (more stable ≈ easier to prepare): always-same user facts → slow-changing histories → ephemeral live state. Unstable sources resist preparability and amplify latency pressure.
Recommended loop: mind-map desired knowledge, list obtainable sources, ship the obvious near ones first, then extend to exotic sources as context engineering matures.
Sources
- Prompt Engineering for LLMs (Berryman), ch. 5