Add ability to choose LLM prompt template for value generation and flow generation

[sklinglernv]

Description

Adds the ability to select a LLM prompt template for value generations and for flow generations by setting a Jiinja variable {% set template = "<name of template here>" %} this allows for greater flexibility and selection LLM prompts based on the intended use case.

This can be used like this

import core

flow generate antonym
  """
  {% set template = "generate_antonym" %}
  bot say "lucky"
  """
  ...

flow main
  when user said "bird"
    $test = ..."a random bird name{{% set template = 'repeat' %}}"
    await bot say $test
  or when user said "antonym"
    generate antonym

with the following config.yaml

colang_version: "2.x"

models:
    - type: main
      engine: openai
      model: gpt-3.5-turbo-instruct

prompts:
    - task: generate_antonym
      models:
          - openai/gpt-3.5-turbo
          - openai/gpt-4
      messages:
          - type: user
            content: |-
                Generate the antonym of the bot expression below. Use the syntax: bot say "<antonym goes here>".
          - type: user
            content: |-
                YOUR TASK:
                {{ flow_nld }}

    - task: repeat
      models:
          - openai/gpt-3.5-turbo
          - openai/gpt-4
      messages:
          - type: system
            content: |
                Your are a value generation bot that needs to generate a value for the ${{ var_name }} variable based on instructions form the user.
                Be very precised and always pick the most suitable variable type (e.g. double quotes for strings). Only generated the value and do not provide any additional response.
          - type: user
            content: |
                {{ instructions }} three times
                Assign the generated value to:
                ${{ var_name }} =

Related Issue(s)

Checklist

• I've read the CONTRIBUTING guidelines.
• I've updated the documentation if applicable.
• I've added tests if applicable.
• @mentions of the person or team responsible for reviewing proposed changes.

[Pouyanpi]

I might be wrong, but maybe

Suggested change

	:param out_variables: If not None the dict will be populated with variables set in the template
	:param out_variables: If not None, the dict will be populated with the variables that were set in the template after rendering. This allows the caller to access the values of the variables used in the rendered template.

[Pouyanpi]

@sklinglernv what do you think of adding a test case for a non-existing template so that we can test behavior when out_variables is missing expected keys in value generation.

For example something like:

        colang_content="""
        flow main
          match UtteranceUserActionFinished(final_transcript="hi")
          $test = ..."a random bird name{{% set template = 'non_existent_template' %}}"
          await UtteranceBotAction(script=$test)
        """

[Pouyanpi]

we were using llm_calls in other tests, using something like

    info = chat.app.explain()
    assert info.llm_calls[0].prompt == expected_prompt

as in test_embeddings_only_user_messages.py which works for Colang 2.0 and in test_general_instructions.py

I think something is not allowing you to use it here, right?

[Pouyanpi]

Thank you @sklinglernv, It looks good 👍🏻 , let me know what you think about my suggestions.