I am struggling on complicated RAG code for Mosaic evaluation deployment

[wind2025]

I want to know how I can apply some complicated RAG designs so I can directly deploy on the Mosaic AI evaluation UI. There are two cases:

1. two or more index retrieval as independent tools (parallel retrievals)

2.two index retrieval processes that depend on each other. For example, Retriever A returns Entity B, which is used as a filter for another Retriever C; C returns the entity D for the LLM invoke function (linear retrievals)

3.two completely independent RAG chains as two tools for the same agent: different retrievals, different LLMs (parallel RAGs)

Can anyone please give me any notebooks or chain templates on any of these case so that the code can successfully pass model.log() and model.deploy()?model.deploy() is very hard to debug on these issues. Thanks!😊

[Mantsama4]

Interesting use case, would love to know the answer!

[dkushari]

Hi @wind2025, since you posted this question, progress has been made quite a bit. You can now use Agent Bricks to construct a multi-agent supervisor tailored to your specific use case. Take a look at this. If you prefer to use code, consider this example.

[CharlotteMarti2]

Hi @wind2025, for Mosaic RAG setups:

Parallel retrievals: independent retrievers feeding one or multiple LLMs.

Linear retrievals: chain retrievers so output of one filters the next.

Independent RAG chains: define separate chains and register both as tools for the same agent.

Test each retriever/LLM with model.log() before model.deploy() to simplify debugging.