question to RAG with LLama Index example

[leviataniac]

Hey guys, we try to implement based on the docling with RAG and llamaindex examples to milvus vector store and we see in milvus the chunked vectors from a n-page pdf document with hashes and so on. thats good. I see as well in thh dl_doc.file_info as well pagereferences and page hashes. Its somehow possible to store the chunked text/vectors with refrenece to a certtain page_hash /number, so we can identifiy on query on which page the text was originally located. how do you think about that useCase?

looking at thats code piece:

            excl_metadata_keys = ["dl_doc_hash"]
            #Here it should be possible somehwo with page refreences???.
            li_doc = LIDocument(
                doc_id=dl_doc.file_info.document_hash,
                text=text,
                excluded_embed_metadata_keys=excl_metadata_keys,
                excluded_llm_metadata_keys=excl_metadata_keys,
            )
            li_doc.metadata = DocumentMetadata(
                dl_doc_hash=dl_doc.file_info.document_hash
            ).model_dump()
            yield li_doc

[vagenas]

Hi @leviataniac, our example notebook has until now showcased using a reference (path) for pointing to the respective part of the original Docling Document (where page nr. etc. is stored and could be retrieved), but it can indeed be helpful to "cache" / inject this info directly into the LlamaIndex metadata.

As shown in this notebook, one can easily introduce this metadata into the chunks (by leveraging a Docling layout-aware chunker).

[vagenas]

We have meanwhile published Docling as an official LlamaIndex integration, check it out here - it also shows how to easily get the page etc.

[leviataniac]

kindly asking, when, do you think would be the DoclingNodeParser v2 available in the official llama_index integration?

[vagenas]

Hi @leviataniac, our PR has now been submitted (run-llama/llama_index#16677) so hopefully very soon! 😉

[leviataniac]

very nice..:)) looking forward.! i got it today with some adjustments and copy paste with hierarchalchunker as well running, but very ugly., so would be cool to have code simpler..:)

[vagenas]

@leviataniac FYI the integration packages for Docling v2 per se have meanwhile been released:

• https://pypi.org/project/llama-index-readers-docling/
• https://pypi.org/project/llama-index-node-parser-docling/

Until the LlamaIndex docs get refreshed (i.e. with next llama-index-core release, ETA: 2-3 days), one can check out the following updated resources:

• Demo notebook
• Docling Reader
  • README
  • Code
• Docling Node Parser
  • README
  • Code

[maurogatti]

@vagenas I am trying to execute your code

https://github.com/DS4SD/docling/blob/9280ce4a8d1a8a6ab6622a93e881bd3f1bcd3d49/examples/rag_llamaindex.ipynb

but I get this error:

`ImportError Traceback (most recent call last)
Cell In[5], line 3
1 from typing import Any, Iterable, Sequence
----> 3 from docling_core.transforms.chunker import ChunkWithMetadata, HierarchicalChunker
4 from docling_core.types import Document as DLDocument
5 from llama_index.core import Document as LIDocument

ImportError: cannot import name 'ChunkWithMetadata' from 'docling_core.transforms.chunker' (/Users/maurogatti/anaconda3/envs/rag-llamaindex-03/lib/python3.11/site-packages/docling_core/transforms/chunker/init.py)`

Are these installation instructions obsolete?

# requirements for this example: %pip install -qq docling docling-core python-dotenv llama-index-embeddings-huggingface llama-index-llms-huggingface-api llama-index-vector-stores-milvus

[vagenas]

@maurogatti the URL you shared refers to an older notebook version (9280ce4), which depends on an older (v1) Docling version.

For the latest version, utilizing the above-discussed updated LlamaIndex extension, please check the docs here (or source here).

[vagenas]

The official LlamaIndex docs have now been updated to contain the changes for Docling v2.
All updated links can be found here.

@leviataniac @maurogatti does this answer your questions so far or are there still open points regarding LlamaIndex?

[leviataniac]

Thank you very much @vagenas. It's perfectly fit. now our code can be simplified with that official extension.

[maurogatti]

Hi @vagenas, I am now able to execute the document conversion. I have many warnings like those below but no blocking error: WARNING: content stream objects stream 182 0 (content, offset 1919): parse error while reading object WARNING: content stream objects stream 183 0 (content, offset 117): treating unexpected array close token as null WARNING: content stream objects stream 183 0, content at offset 0: operation for array attempted on object of type string: treating as empty Thanks a lot