Hybrid ML-SLM RAG System for Large Technical PDFs

We present a novel hybrid Retrieval-Augmented Generation (RAG) system designed specifically for processing and extracting information from large technical PDFs. Our system addresses the challenge of efficiently managing extensive technical documentation by combining traditional machine learning approaches with Small language models. The architecture employs a dual-retrieval mechanism that leverages both TF-IDF vectorization and dense neural embeddings to optimize information retrieval accuracy from complex technical content. The system features an advanced PDF processor with technical content-aware chunking strategies that preserve the integrity of specialized content such as mathematical formulas and section hierarchies. Memory efficiency is achieved through intelligent batch processing and optimized vector storage using FAISS when available. For generation, we integrate Google’s Gemma 3 1b got Highest accuracy 68+ and other SLM with failover mechanisms to ensure consistent output quality. Empirical evaluation demonstrates that our hybrid approach outperforms single-method retrieval systems when handling specialized technical terminology and complex conceptual relationships. The system successfully processes collections of 8-15 large PDFs (500+ pages each) while maintaining contextual relevance in responses. This approach represents a significant advancement in making large technical document collections more accessible and queryable through natural language, with particular relevance for scientific research, technical documentation management, and educational Books applications.