Enhance prompting with techniques heard at Anthropic Vertex AI RAG talk

[MrCsabaToth]

Related repo: https://github.com/anthropics/courses/tree/master/real_world_prompting

There were some good blueprints about prompt sections and the ordering of those:

Note about the 2nd section: "Dynamic content / RAG (step 2) also covers Tools listing? How about chat history? What should be the order of RAG doc snippets, tool / function listing, tool return results, chat history?" Answer:

• Tool definitions: kinda with the system prompt, up front
• If there are any tools examples that would go to the example section (below, later)
• RAG doc snippets: in the dynamic section
• Chat history: in the example section

More slides:

Related: #44

[MrCsabaToth]

Note that Anthropic prompt conversations can differ from Gemini or OpenAI or other LLMs, it mainly depends on how they were trained and RLHFd.
We may need to make small adjustments but so far Gemini is able to understand the XML tagging, to the point that it only responds with the closing tag because I prefilled the prompt with the opening tag.

[MrCsabaToth]

We mostly refactored the prompts. It's a question how much these Anthropic style XML tags are bode well with Gemini. Time will tell.