[Multi-hop RAG opt + improvement] [V0.3.0 Cycle Graph]

[liyin2015]

• fix the api in the word count code and generate better diagnose report
• multi-hop RAG will have cycles, we can create a copy to every node that is a cycle to unroll the DCG to DAG, but hard to do as we have to recreate the component such as Retriever or Generator. DsPY simply does not support cycle yet.

Design choice: (1) support cycle, we treat it as a component where its prompt has to be adapted to multiple sets of inputs and outputs, in practical this make sense (2) user not creating a cycle, they treating each repeating as a new copy node and has its own parameter.

For (2) its easy, everything works out of box, users just need to know that they are doing this intentional
for (1) we have to carefully design how the gradient works, and how the demo works (demo should aggregate both parts)

• (2) User try not to create a cycle

Multi-hop retriever

Multi-hop RAG (works by simplify replace the retriever if we simplify follow the retriever api)

[sameeramin]

Hey @liyin2015, Is this PR ready to be reviewed?

[liyin2015]

We can diagnose and trace all generator calls end to end as follows:

It can track two generators used in the retriever which is saved in ComponentList but as each file trace each call separately, we might want a single file that will have better visualization, especially fit it back to a graph.

[liyin2015]

Show data id (from the dataset) in the graph, and show the score for each traces in the demo paramter.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[Sylph-AI]

TODO: write a better relase change log later