Metadata extracted and Contextual Retrieval

[volethanh]

Hi, this framework is amazing and I try to apply to my RAG app and the result is great. But I see there is some thing that needs to improve.

It is the extracted json metadata, because when User upload document in the backend admin, at this process, It will generate metadata of this document (include file level and chunk level), then insert to vectordb.

But back to chat app, this will be a different frontend chat app for end user and at this time, when user send the query, it won’t get metadata info that is extracted above, of course we can store this metadata json in db or somewhere else, but may be the schema of metadata will not consistently when uploading another documents.

Do you have any ideas about this?

Another things is that, they have just publish Contextual Retrieval, I think it would be the same idea with this framework, what do you think about this ?

https://github.com/run-llama/llama_index/blob/main/docs/docs/examples/cookbooks/contextual_retrieval.ipynb

Thanks for creating such an amazing framework

[darshil3011]

Hi @volethanh thank you for your kind words. Glad that its helping you !

I understand your concern regarding metadata for uploading documents. Please understand that this framework, although is dynamic, is limited to serve similar type of documents at a time. For eg. you cannot use it for medical documents and legal documents because both require different metadata schema. However, it is very easy to use this framework out of the box (as-is) for both use cases separately, which is the beauty of this framework. It avoids all the heavy-lifting required to setup RAG and get accurate results for your use case.

I would love to know if anyone in the community has any ideas on making this even more dynamic by allowing multi-domain documents simultaneously.

I will check contextual retrieval and see if we can add support for it to our existing framework.

We are adding support for more vectordb and other LLMs using Nvidia NIMs soon ! Stay tuned and thanks for all the support