AZØR TRAINING

• gemini-CLI
  • download Gemini CLI: https://github.com/google-gemini/gemini-cli
  • set up Gemini API key: https://aistudio.google.com/app/api-keys
• open router: https://openrouter.ai/workspaces/
• openAI: https://platform.openai.com/settings/organization/billing/overview
• SDKs:
  • python (download platform): https://www.python.org/downloads/
  • nodejs (download platform): https://nodejs.org/en/download

Question

(!) What are the patterns / tricks to overcome limitations of the context window?

• short-term memory vs long-term memory?
• who needs planning (such as in cursor, claude code, etc)
• how do commands work (cursor, claude code)
• claude code uses codebase indexing / vector dbs. Claude Code doesn't. What's the tradeoff behind it?
• how subagents works and how to use them efficiently
• what tooling is being used when running a model locally?
• how does an AI API/SDK know where to send requests?
• how to make a jailbreak?

Design/Implementation exercises

• [IMPL] SUPPORT PARAMETERS (top p, top k, temperature) for client APIs used in Azor
• [DESIGN/IMPL] MORE ASSISTANTS / CHANGE THE ASSISTANT
  • the session should have information about the selected assistant (the assistant is defined per session)
    • when saving a session, not only the conversation history is saved, but also the assistant
    • when loading an old session – similarly
    • changing the assistant during a session leaves a trace in the conversation history so that the model knows in subsequent steps that a change has occurred (then, based on the prompt system, it will better understand the context)
• [DESIGN/IMPL] PROVIDE SESSION TITLE:
  • You should give the conversation a TITLE, so that when you search for a conversation from the list, you can see the title in addition to the ID.
    • Design a solution - who, when, and how determines the title?
    • Minimalist variant - a separate shot at the model at the beginning of the thread...
    • Do your research and design a solution to see how the "minimalist" variant can be improved.
  • persistence: the conversation title is stored and saved in the session (and the session file).
  • The title can be changed (e.g., based on the new command /session rename NEW_TITLE)
  • It can also be previewed (e.g., based on the new command /session title -> EXISTING_TITLE)
• [DESIGN/IMPL] CLARIFY QUESTION/TASK
  • gives the model the ability to ask the user when something is unclear:
    • the user wants something from the model
    • the model responds with a question, giving, for example, 4 options to choose from (like perplexity)
    • the user selects one of the options
    • the model finally responds
• [DESIGN] AUTONOMUOUS CONVERSATION BETWEEN 2 ASSISTANTS
  • initially human sets the topic via prompt
  • human interaction is done
  • assistant 1 answers to human prompt
  • assistant 2 answers to assistant 1
  • assistant 1 answers to assistant 2
  • assistant 2 answers to assistant 1
  • assistant 1 answers to assistant 2
  • ...
  • human can only close the conversation somehow (SIGINT, whatever)