How do I enhance my PDF RAG App's mathematical capabilities ?

[green steppe]

Hello everyone,
I'm currently working on a multimodal PDF RAG app ( to do QA with PDFs containing texts, images, tables ) .
The core of it is a RAG chain which takes the user query and returns the answer. It works for text , returns images and able to display the tables and answers from it .
When I ask math related questions from the tables in the pdf , it fails badly.
Currently I've modified my system prompt asking the LLM to double check , perform calculations in step by step manner etc., still I don't get correct answers .
Mathematical Operations Format:
Step 1: Define the objective
Step 2: List source data with references
Step 3: Show the calculation setup
Step 4: Perform step-by-step operations
Step 5: Verify results
Step 6: Present the final result with context
above is the snippet from my system prompt. Is this enough ?
What can I do to enhance my app's mathematical capabilities ?
Should I use an agent instead of a normal LCEL chain ?

[solid sparrow]

Have you read through the responses and understood whether the issues are primarily coming from a) incorrect extraction from the tables or b) incorrect arithmetic?

Or c) the RAG itself not picking up the right tables