I don't understand how the "Router" works

[tianhaolin1991]

According to your introduction, main.py reads data from "./Loong/data/loong_process.jsonl" (which is not disclosed in the project).
It seems that you use the 'Router' in the structured data generating process.
And then the data is used as the context(or search results) to help the LLM answer your question in the subsequent process.
But in common scenarios, we always have one or more knowledge Bases in our system, and we can not give several titles of doc as a parameter of the method 'do_route' because we can't get all relevant docs together actually.
So in my opinion, how Struct-Rag can be used as a plugin in the enterprise-level RAG System is a problem. So do you have any good suggestions?

[Li-Z-Q]

Thank you for your attention. The question you raised is very important.

In fact, our work does not focus on large-scale knowledge bases. The experiments considered only small-scale, relatively clean knowledge bases (with each document having a preprocessed title).

Regarding the enterprise-level RAG system you mentioned, I believe its main challenge goes beyond the scope of our work, and the difficulty lies in the retrieval process. For large-scale knowledge bases, I think it's essential to have a retriever that fetches a small set of relevant documents, then uses an LLM or rule-based methods to extract the core content of these documents. Only then can our work come into play, which involves using a Router to determine the optimal structured type and then constructing the information into structured knowledge to enhance the LLM.

If you have more you'd like to discuss, feel free to reach out anytime.