[Issue]: Agentic Loop over a large document

[ismailsimsek]

Describe the issue

Is it possible to read large PDF document in chunks using agents. Without programmatic loop.

Could it be done using task-decomposition? have anyone done something similar?

Steps to reproduce

something like below:

1. Agents uses tool and reads PDF document 50 pages each time (total 300 pages)
2. Agents summarizes all the chunks to 6 page output
3. Summary written to file

Screenshots and logs

No response

Additional Information

No response

[WaelKarkoub]

@ismailsimsek are you asking for an OCR capability or a rag capability? If OCR, I believe it's planned in the multimodality road map #1975

[ismailsimsek]

Currently trying to get it work with RAG ( RetrieveUserProxyAgent + GroupChatManager )

Appreciate if anyone could point to similar solutions..

Current code is here: https://github.com/ismailsimsek/aistorybooks/blob/story-book/classic_storiesv2.py
PR ismailsimsek/aistorybooks#3

currently just trying to summarize PDF, later on planning to add image generation too

[thinkall]

Currently trying to get it work with RAG ( RetrieveUserProxyAgent + GroupChatManager )

Appreciate if anyone could point to similar solutions..

Current code is here: https://github.com/ismailsimsek/aistorybooks/blob/story-book/classic_storiesv2.py PR ismailsimsek/aistorybooks#3

currently just trying to summarize PDF, later on planning to add image generation too

The current RetrieveUserProxyAgent should support PDF files. Have you tried it?

[ismailsimsek]

@thinkall i will check it. what i am looking into is summarizing the PDF in small chunks, since its too big. in a loop, is that possible using the agents to loop and process chunks one by one?

[thinkall]

@thinkall i will check it. what i am looking into is summarizing the PDF in small chunks, since its too big. in a loop, is that possible using the agents to loop and process chunks one by one?

The agent will split the pdf into chunks and save it into vector db.

[thinkall]

Close as it's not active for a long time. Please reopen if the issue still persist.