Demo v2: Auto-correct JSON format via loop

[tholor]

Motivation

Pipelines in Haystack 2.0 are more flexible and more robust than ever.
They allow you now to run loops within your pipeline. Let's look at why this matters.

Exemplary Use Case: Validate & auto-correct the output format of an LLM

Let's imagine you want to extract structured information from a document in JSON format.
As your further processing steps depend on the correctness of this JSON format, you want to make sure that the LLM really produces it in a robust way.

Approach

We validate the output of the LLM against a pydantic schema. If the output is valid, we return it to the user. If not, we create another prompt that leverages the generated output + error message. To avoid infinite loops we limit it to max. 5 iterations.

Demo: Colab

To Dos:

• We need optional variables in PromptTemplates (we are using a haystack branch so far)
• Allow three variables in the prompt template so that we can include the error message (so far only the wrongly generated output is used)

[ZanSara]

Left two small fixes that make this work on Canals 0.10

[bilgeyucel]

Closing this one as we have the example as a tutorial: https://github.com/deepset-ai/haystack-tutorials/blob/main/tutorials/28_Structured_Output_With_Loop.ipynb