Give examples on how more complex pipelines should be build

[wirtsi]

Hey ...

so for example, looking at the Function Calling Example, you notice that the whole looping (so feeding the RAG documents back into the ChatMessages list) is happening outside of the pipeline (the ChatGenerator is called over and over again).

This does not seem very close to how Pipelines are supposed to be used, I am under the impression that the idea is to build a directed graph (DAG) that haystack executes. So something like

PromptBuilder
OpenAIChatGenerator
FunctionCallDetector
if true
> FunctionCaller
> Loop reply back to PromptBuilder by appending to messages
if false
> return output

So that the whole execution handling resides inside the pipeline. It would also be interesting to see how error handling would be done. I think it would be nice to have a final pipeline state which either gives me the response or an error message so that my code can work with this.

So generally some more advances examples how to conditionally execute pipelines and potentially also how to break down pipelines into sub-pipelines

Bests & Thanks

[julian-risch]

Hey @wirtsi you're right, it shouldn't be the case that the whole looping has to stay outside the pipeline. With Haystack, you can also have connections of components in a pipeline that create a loop. In that case, the directed acyclic graph (DAG) actually becomes a directed cyclic graph. We have recently added an example showing it for a similar use case: conversational rag using experimental memory components: https://colab.research.google.com/github/deepset-ai/haystack-cookbook/blob/main/notebooks/conversational_rag_using_memory.ipynb
For checking the outputs of a pipeline, maybe you haven't seen our include_outputs_from parameter. It's described in the docs here: https://docs.haystack.deepset.ai/docs/debugging-pipelines and you might find it helpful.
Thank you so much for the feedback that advanced "examples how to conditionally execute pipelines and potentially also how to break down pipelines into sub-pipelines" would be helpful! 🙏

[wirtsi]

Hey @julian-risch, thanks for your response. I was not aware of the example that is using the memory, this is really helpful.

I managed to apply the examples given into my code base and got it to work ... but there are still some parts of this whole DAG/DCG pipeline approach that I don't fully get. Like one thing that baffles me:

How do you work with pipeline components that receive no input. Like with the ChatMessageReceiver ... take this graph for example

• I start by pushing the system prompt and the user prompt into the joiner at the bottom.
• The memory writer persists them
• I move to the loop counter (which accepts any input and outputs a boolean max_reached)
• The router routes on max_reached, let's say this is false
• I try to output None from the router on the regular_router but since memory_reader has no defined input, the pipeline fails here

I can of course override the ChatMessageRetriever to accept any input, but it rather feels I am missing a part in the understanding

Thanks in advance

wirtsi