Development of an Intelligent Conversational Assistant for Semantic Exploration of Scientific Articles

Scientific publishing now grows faster than researchers can track it, and keyword search doesn't help much with that problem. It matches text, not meaning, so when a researcher is exploring a topic rather than hunting a known phrase, keyword tools tend to miss what they're actually after. This paper describes a conversational assistant built to search arXiv by meaning instead of keywords. It pulls metadata through the arXiv API, cleans the text, and generates sentence embeddings with a pretrained Sentence Transformer (all-MiniLM-L6-v2). Those embeddings feed a FAISS index (IndexIVFPQ), which ranks documents by cosine similarity, and a Streamlit interface wraps the whole thing in a conversation. The corpus covers more than 30,000 articles from 2010 to 2024, spanning computer science, physics, mathematics, and biology. On a set of test queries the system hit 92% Precision@5 and an intra-cluster similarity of 0.89, with search latency under 50 milliseconds and a memory footprint around 200 MB. It held up at one million vectors without degrading. A controlled comparison against a BM25 baseline backs up the core claim: pairing semantic embeddings with an approximate nearest-neighbor index and a conversational interface retrieves literature better than keyword search does, with Precision@5 and Precision@10 improvements significant at p<0.01. The system still trips on short, ambiguous queries, has no personalization, and only draws from one source. Reranking, citation-graph reasoning, and pulling in additional sources are the next steps.