I m working on a legal document It is

extra_info is definitely used by the index

I'm not sure what you mean by atomicity, like you want to maintain the order?

If I ask a question, I want to find all the legal articles that are related to this question.

A article is in a section which is in a chapter. I want to keep this information also.

Thks for your help

Hmm. If you create your documents as each document is an article, you can store the section and chapter in the extra_info

document = Document('article text', extra_info={'section': section, 'chapter': chapter})

That what I was thinking.
But I'm not sure all the index will use this information. Is extra_info attached to each node ?

I was also thinking to use compostable graph. One index by chapter and put the structure ( chapter 1/ section 1 / paragraph 1 / article 1, 2, 3...) in the summary

Yea the extra-info is attached to each node. So if a document is parsed into 3 nodes, each node will have the same extra info

Then, the response.source_nodes will also show this extra info

We are planning to use local model like vicuna. Also the text is in French.

Do you have advice? What would be for you the best approach with the index ?

local model should be fine! We have a HuggingfaceLLMPredictor which might help a lot (it can be weird to setup, but happy to help if you get stuck)
https://gpt-index.readthedocs.io/en/latest/how_to/customization/custom_llms.html#example-using-a-huggingface-llm

A simple vector index and splitting the text into articles is probably the best approach actually 🙂 If the results aren't great, there is some things to tweak like top k or chunk size. If you move to a graph, the response time will be slower