Implementation of an Interactive Query System Using Nomic Text Embed, DeepSeek R1 1.5B, and Cosine Similarity rankers

Our work in this research builds and demonstrates an interactive query system using Retrieval Augmented Generation (RAG) to generate accurate responses. For evaluation, the system uses Nomic Text Embed and ChromaDB, and is evaluated on the rag-mini-wikipedia dataset from Hugging Face. The cosine similarity re-ranking improves retrieval precision. Long context query (up to 128,000 tokens) can be handled by DeepSeek R1 1.5B and coherence and reasoning can be improved. Chain of Thought (CoT) and Chain of Notes prompt engineering techniques improve factual accuracy of this work, reduce hallucinations and enhance multi-hop reasoning. The evaluation follows LangSmith framework in terms of accuracy. Through this study we show that combining advanced embeddings, reranking, and LLMs improves the main performance of the query in AI driven knowledge retrieval. Our approach achieves 73% accuracy, surpassing both the vanilla BERT model (59.68% accuracy) and the BERT (SQuAD) model (72.36% accuracy) reported in related work of march 2024 . These improvements reflect the effectiveness of combining advanced embeddings, re-ranking, and a locally hosted LLM for AI-driven knowledge retrieval. The evaluation methodology follows the LangSmith framework, highlighting how robust retrieval and carefully engineered prompts can boost system performance on complex QA tasks.