Merge Router Agent and Summarizer Agent

[CakeCrusher]

Objective

Utilize DSPy to develop a pipeline that programmatically optimizes instructions for the following task.
Experiment merging the functionality of #1 (Router Agent) #3 (Summarizer Agent) to generate both the transition request and the generated summary.
The resulting instruction should be concise (aim for <500 tokens) with high consistency of Transition string generation (aim for >95%) and decent decision accuracy (aim for >70%). These qualifiers should be proven via DSPy.
Instead of this:

It should flow like so:

Input

(not sufficiently knowledgable on DSPy to comment)

Output

(not sufficiently knowledgable on DSPy to comment)

Example

Prompt:
(Chat history provided in completions API that could be 1000s of tokens long)
Instruction:

Your role is to determine of tools are needed or if a normal response is sufficient. If tools are necessary then create summarization to be used during tool use to accomplish the current user request, also ensure that it contains the word TRANSITION. Otherwise just respond normally.

The tools available to you are:
- Text Generation:  general text response
- Retrieval: retrieve information related to the last message

Use 1
Chat History:
...
I like writing about llamas
...
User Message:
What did I write about

Agent Response:
You are to TRANSITION. The user likes to write about llamas and now wants to know what he wrote, you may want to use retrieval for this task ...
Use 2
User Message:
When did ww2 happen.

Agent Response:
World War 2 began in 1941 ...

Test

Testing is core to this task but will already be done in the process.

Resources & Suggestions

• May want to leverage synthetic data for test cases. Review of synthetic data: https://eugeneyan.com/writing/synthetic/
• DSPy: https://github.com/stanfordnlp/dspy
• Article on DSPy: https://bassimeledath.substack.com/p/demystifying-dspy-the-framework-redefining