AI Engineering from Scratch（英文原版）
Given a video duration, query complexity (single event vs holistic summary), and open vs closed constraints, pick a long-video strategy and emit a config.
Produce:
- Strategy pick. Brute-context, ring-attention (LongVILA), token-compression (Video-XL), or agentic-retrieval (VideoAgent).
- Token budget. Duration * FPS * per-frame-tokens. Warn if > LLM context.
- Expected recall. Needle-in-a-haystack recall at video-length percentiles. Cite Gemini 1.5 reports when relevant.
- Latency. Prefill time for brute-context; retrieval + VLM for agentic.
- Engineering path. Code snippet scaffold for the chosen strategy.
- Fallback plan. Hybrid: brute-context global summary + agentic local detail.
Hard rejects:
- Proposing brute-context for a 2-hour video on an open 72B model. Context does not fit.
- Claiming agentic retrieval always wins. For holistic-summary questions it loses to brute context.
- Recommending token compression without flagging the recall tax.
Refusal rules:
- If target is a 90-minute video at frontier recall (>95%), refuse open-only options and recommend Gemini 2.5 Pro.
- If user cannot afford tool-calling loops, refuse agentic-retrieval and propose compressed brute-context.
- If user needs real-time (stream-as-it-plays), refuse retrieval (too slow) and recommend streaming Qwen2.5-VL.
Output: one-page plan with strategy, budget, recall, latency, engineering path, and fallback. End with arXiv 2403.05530 (Gemini 1.5) and 2403.10517 (VideoAgent) for comparison.