I am using Assistants, RAG (via assistant) and few shot examples for inspiration (via the assistant instructions) through the API. I’ve created a thread and am looping through 16 topics, creating a new message + run for each category.
Right now, the bulk of my logic (Overall instructions + RAG + Chain of Thought Instructions + few shot examples) is in the Assistant, but I find myself really struggling with where to put examples, where to enforce file retrieval, etc.
Also, I’d like to retain at least a bit of the context per category so that’s why I’m creating a new thread message for each category. Effectively, it kinda looks like this:
assistant_instructions= f"{overview_text} + {step_by_step_instructions} + {chain_of_thought_instructions} + {few_examples}"
my_assistant = client.beta.assistants.create(name="my_assistant", instructions=assistant_instructions, tools=[{"type": "file_retrieval"}], model="gpt-4o")
thread = client.beta.threads.create()
for i, topic in enumerate(topics):
# this next bit alludes to the RAG reference docs
thread_content = f"use the information in the examples on topic {topic['name']} as inspiration, and generate 4 reasonable prompt/completion pairs that adhere to the design in {topic["design"]}"
message = client.beta.threads.messages.create(thread_id = thread.id, role="user", content=thread_content)
run = client.beta.threads.run.create_and_poll(thread_id = thread.id, run_id = i, assistant_id = my_assistant.id)
However, I can’t shake the feeling that maybe I shouldn’t be creating a new run AND thread message for each topic, AND I can’t shake the idea that maybe I should be doing something different with the examples. Can anyone please share a bit of guidance on how to structure such a thing?