Hi TensorZero team,

I'm working on migrating a complex prompt chain from our codebase to TensorZero. Our current implementation uses a multi-turn conversation format with system message, followed by 4 user/assistant message pairs as few-shot examples, and then a final user message with dynamic content.

Currently, it seems TensorZero's function/variant system is designed around single request-response pairs. I'm wondering:

1. Is there native support for multi-turn conversation templates? For example, being able to define a sequence of user/assistant templates within a single function variant.

2. If not, are there recommended patterns for this use case? Some possibilities I've considered:

   • Concatenating all examples into a single system/user prompt (loses the conversational structure)
   • Chaining multiple functions together (is this possible?)

3. Is this a feature you'd consider adding? Multi-turn conversations are common in few-shot prompting scenarios, especially for tasks like:

   • In-context learning with examples
   • Chain-of-thought reasoning
   • Task demonstrations

Here's a simplified example of what I'm trying to achieve:

[functions.my_function.variants.my_variant]
type = "chat_completion"
model = "openai::gpt-4"
# Ideally something like:
messages = [
  { role = "system", template = "system.minijinja" },
  { role = "user", template = "example1_user.minijinja" },
  { role = "assistant", template = "example1_assistant.minijinja" },
  { role = "user", template = "example2_user.minijinja" },
  { role = "assistant", template = "example2_assistant.minijinja" },
  # ... more examples ...
  { role = "user", template = "final_user.minijinja", use_schema = true }
]

Would love to hear your thoughts on the best way to handle this pattern within TensorZero's architecture.

Thanks for building such a thoughtful tool!