# Manage Context Windows in an Applied AI System Design the context-building layer for an applied AI assistant. It must combine the current request, conversation history, retrieved evidence, and tool results while staying within a model's context limit and keeping important evidence traceable. ### Constraints & Assumptions - The model has a fixed token budget shared by input and expected output. - Retrieved documents can conflict, repeat one another, or contain untrusted instructions. - Conversation history can grow indefinitely. - The system must explain which evidence supported the answer. ### Clarifying Questions to Ask - Which information is authoritative, and which is merely conversational context? - What output-token reserve and latency target are required? - Can old conversation turns be summarized, and must summaries be reversible to source turns? ### Part 1 — Context construction Propose a budget and ordering policy for system instructions, user request, history, retrieved evidence, and tool results. #### What This Part Should Cover - A token budget with output reserve - Priority and truncation rules - Deduplication and provenance ### Part 2 — Safety and quality Handle prompt injection, conflicting evidence, stale summaries, and evaluation of context selection. #### What This Part Should Cover - Trust boundaries and instruction hierarchy - Evidence recency and conflict reporting - Retrieval and answer-quality measurements ### What a Strong Answer Covers - A deterministic context policy - Traceable evidence and summaries - Evaluation that separates retrieval from generation ```hint Budget by information role Do not concatenate until the limit is reached. Reserve explicit budgets for instructions, user intent, evidence, history, and output, then choose content within each role. ``` ### Follow-up Questions - When should a larger-context model be chosen? - How would you detect that summarization removed a critical constraint?