Quiz dotyczący zagęszczania kontekstu i zarządzania historią rozmów

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. Dlaczego długie sesje czatu stają się coraz droższe?
   • The model stores a copy of the chat on every server
   • Providers bill by the number of days a chat stays open
   • Most chat APIs are stateless, so every request resends the full conversation so far
   • The model retrains itself on each new message
2. Jaka jest główna słabość obcinania okna przesuwnego?
   • It forgets early facts, such as constraints the user stated at the start
   • It is the most expensive strategy
   • It requires a second model to run
   • It rewrites the system prompt
3. Które ryzyko jest specyficzne dla podsumowania kroczącego?
   • Summaries cannot be generated by language models
   • It increases the number of tokens per request
   • It only works with image inputs
   • Summaries lose detail, and errors can compound when a summary is summarized again
4. Co zwykle powoduje zagęszczanie w narzędziach agentowych?
   • An empty history with no information carried over
   • A structured handoff of goal, progress, decisions and open problems that seeds a fresh context
   • A full copy of every file the agent has opened
   • A newly fine-tuned model
5. Dlaczego czyszczenie wyników starych narzędzi jest tak skuteczne w sesjach agentów?
   • Large outputs from searches, file reads and commands are often the biggest token consumers and are rarely needed again
   • Tool results are always incorrect
   • APIs forbid tool results after the first turn
   • It makes the tokenizer run faster
6. Why should thresholds be based on token counts rather than message counts?
   • Messages cannot be counted programmatically
   • Tokens are free, while messages are billed
   • A single tool result can outweigh many chat turns
   • Providers bill only by message count
7. When clearing a tool result, why replace its content instead of deleting the message?
   • To make the history longer
   • Many APIs reject a history in which a tool call has no matching result
   • Placeholders make the model more creative
   • Replacing content encrypts it
8. According to the guide, where does continuity in a long session actually come from?
   • The model's permanent memory of all past chats
   • The GPU's internal cache
   • The user's web browser
   • What the application chooses to resend or retrieve