AI-Driven RFP Document Intelligence and Question-Answering Pipeline

Abstract

This paper presents an AI-driven pipeline for intelligent ingestion and question-answering over Request for Proposal (RFP) documents. RFP documents are typically 50 to 200+ pages in length, making manual information retrieval tedious and error-prone. We propose a Retrieval-Augmented Generation (RAG) architecture that automates extraction, chunking, embedding, and context-grounded answer generation. The system ingests RFP documents via IMAP email polling or manual upload, applies a three-tier PDF parsing strategy (pdfplumber, Unstructured, and OCR fallback), and stores 384-dimensional semantic embeddings in a Pinecone serverless vector database. User queries are semantically matched to the top-K=5 most relevant document chunks using cosine similarity, and these chunks are injected into a grounded prompt for Llama 3 (via Ollama) running entirely on local hardware. Token-by-token responses are streamed to the frontend dashboard using Server-Sent Events (SSE). Experimental results show retrieval similarity scores ranging from 0.75 to 0.89, with zero hallucination due to strict context grounding. The system supports multi-document comparison and ensures complete data sovereignty by keeping all inference local. This work demonstrates significant efficiency gains in RFP analysis for procurement and tendering workflows.