How to implement a bot that summarizes given statements and then gives references?

[ifsheldon]

The problem setting is: I have N statements. I want LLM to generate M different insights based on these statements and to give references.

I read the README and it seems role-based-chat-model-example is the most relevant. And the tools we need include geneach, select and each.

My pseudocode would be like

1. List all statements:
   This is easy, like

   ---{{#each statements}}
   Statement {{@index}}: {{this}}{{/each}}
   ---
   

2. Prepare a list of insights. Show the insights in the list to LLM, so that it will not repeat itself.
3. Ask LLM for M times. For each time:
   1. it will see the statements and previous insights.
   2. it will generate a new insight based on the statements
   3. it gives references to the statements that are related to the new insight. BUT how can we ensure that the selected statement indices are valid??

I can for sure call guidance multiple times to do this with proper string formatting, but can we do it in just one call of guidance with one prompt?

An example output would look like (suppose N=3, M=2):

Statement 0: statement.........
Statement 1: statement.........
Statement 2: statement.........

Given the above statements, give me 2 insights. Do not repeat insights and list references.

Insight 0: insight..............
References: [0, 1]

Insight 1: insight..............
References: [1, 2]