Best AI Architecture for Processing and Querying Large PDFs (7000 Pages) with Fast Response Time

Hi everyone,

I’m building an AI-powered educational assistant that must answer questions grounded in very large PDFs (up to ~4000 pages per document). I’m currently using a RAG-based setup, but I’m facing serious production issues and would appreciate architectural guidance.

Current Problems:

1. Very high latency

   • Responses take 25+ seconds.

   • Sometimes even longer with complex queries.

2. Missing information in responses

   • The system retrieves only partial sections.

   • Important parts of the document are ignored.

   • Answers feel incomplete or fragmented.

3. Requirements:

• Fast response time (<3 seconds ideally)

• High-quality, well-structured answers

• Accurate grounding with page references

• Ability to handle 4000+ pages reliably

• Production-ready and scalable

Any one have a better Ai pipeline, or knew how to implement it in a proper way ?