This system uses an agentic retrieval-augmented generation pipeline for academic paper question answering. It combines PDF processing, vector retrieval, LLM-based planning, query rewriting, evidence evaluation, grounded answer generation, conversation memory, and functional evaluation.
The user uploads academic PDF papers and asks a research question.
The system extracts text page by page, splits it into overlapping chunks, and stores page metadata.
Each chunk is embedded and stored in ChromaDB for semantic retrieval.
The planner identifies the task as document-based question answering and activates the agentic pipeline.
The user question is rewritten into a focused retrieval query using important keywords.
The system searches ChromaDB and retrieves the most relevant document chunks.
The retrieved chunks are checked to determine whether they are sufficient to answer the question.
The LLM generates a grounded answer using only retrieved evidence with source, page, and chunk citations.
The system stores the last five interactions in session memory for multi-turn context tracking.
The system compares baseline LLM answers against Agentic RAG answers and scores them using evaluation metrics.
The system is agentic because it does not simply send a question directly to an LLM. Instead, it follows a multi-step pipeline where different agent roles perform planning, query rewriting, retrieval, evidence checking, answer generation, memory tracking, and evaluation.
©2026, System Developed by Abdullah Al Amin