Quiz dotyczący inżynierii kontekstu

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. What best distinguishes context engineering from prompt engineering?
   • It only concerns the wording of the user's question
   • It covers everything the model sees, including retrieved data, tool results and memory, not just the wording
   • It requires fine-tuning the model's weights
   • It applies only to image-generation models
2. What did the 'lost in the middle' research suggest about long inputs?
   • Models read every part of a long input equally well
   • Models ignore the beginning of their input
   • Models often use information at the start and end of a long input better than information in the middle
   • Models cannot process more than about 1,000 tokens
3. Why does the guide describe context as a budget?
   • Every token costs money and latency, and irrelevant material can distract the model
   • Providers charge a flat monthly fee for context
   • Models refuse inputs above a certain dollar amount
   • The word refers only to the compute used in training
4. What is just-in-time loading?
   • Loading the entire codebase into the window at the start of a session
   • Retraining the model whenever new data arrives
   • Caching every model output permanently
   • Giving the agent lightweight references, such as file names, and letting it fetch full content with tools when needed
5. Why is it useful to put stable content, such as system instructions and tool definitions, at the start in a fixed order?
   • Dzięki temu model jest bardziej kreatywny
   • Every API requires this order
   • An identical prefix can be reused through prompt caching, which lowers cost and latency
   • It hides the instructions from the user
6. Which approach to tool design does the guide recommend?
   • Tools should return all raw data so nothing is missed
   • Tools should return concise, high-signal output and support filtering or pagination
   • Many tools with overlapping descriptions give the model more choice
   • Tool descriptions should be left empty to save tokens
7. Which statement is a misconception the guide warns about?
   • Large context windows increase cost per request
   • Tool definitions take up tokens on every turn
   • Retrieval can reduce how much text goes into the window
   • A bigger context window removes the need to curate what goes in
8. What role do sub-agents play in context engineering?
   • They explore in their own separate context and return condensed findings to the main agent
   • They double the context window of the main agent
   • They replace the system prompt
   • They store user passwords between sessions