Whenever you're ready, walk me through how you'd approach this.
Explain how you would implement dynamic context window management for an agentic application handling long-document Q&A.
Discuss how you would select, compress, retrieve, summarize, and replace context during multi-step reasoning. Address token budgeting, latency and cost tradeoffs, answer grounding, citation preservation, prompt injection in documents, and behavior when relevant evidence cannot fit in the available context.
Describe the evaluation strategy and provide implementation-level details.