Learning Steps - Akash / Srijan - Take Vihang’s Help to Find Best Courses

1. Extraction Courses - 2
2. Chunking Courses - 2
3. Fine Tuning Models - 2

Breaking Down Question - Srijan - 3

1. Understand what questions and understand what are keywords in questions
2. Determine Dynamic Chunk size based on query
3. Fetch Number of Chunks Dynamically based on Keywords Found

Retrieval - Srijan - 4

1. Find Best 3 Documents that Can Answer the Question
2. Find Best Pages Which May Have the Answer - 3 Pages From Each Document
3. Find Best Chunks Based on Keywords - Expand Context if Needed
4. Find Best Chunks Based on Suggested Questions - Expand Context If Needed

Extractions - Akash - 4

1. Multimodal Extractions of Pages - Try Extraction in a structured way
2. Try LLMShrepa For Extraction - Test Quality of Extraction
3. Heading and Sub Headings and Maping Paragraphs with it.
4. Maping Paragraphs with keywords

Augmentation - Srijan - 2

1. Trying Nvidia Embed Model for Task Based Ranking of Information
2. Trying Task Based Models for Reranking Content or By Passing through round of LLM

Generation - Akash - 2

1. Generate Answer Based on Page from Multimodal (Try Multi Page Generation)
2. Use a LLM to Combine and generate a Final Answer

Building Testing Setup - Srijan

1. Find 20 Documents of Different Kinds

2. 100 Questions

3. Make a clean new Environment and Populate it with it.

4. Steps →

   1. Creating Clean Environment and Setup of Test Data - Srijan
   2. Query Understanding - Srijan
   3. Trying Retrieval Using Suggested Questions - Srijan
   4. Trying Retrieval Using Keywords - Srijan
   5. Dynamic Chunking and Dynamic Chunk Size Retrieval Based on Query - Srijan
   6. Multimodal Extraction - Akash
   7. LLM Sherpa Extraction - Akash
   8. Multimodal Answering - Akash
   9. Trying Task Based Embeddings Models - Srijan