Vector Search in Large Language Models: Experimental Evaluation
with MongoDB Atlas
Deepak1,∗, Savita Sheoran1
1Department of Computer Science & Engineering, Indira Gandhi University, Meerpur Rewari, India
Emails: dpkrao91@gmail.com; savita.sheoran@igu.ac.in
Abstract
The growth of Large Language Models (LLMs) applications has intensified the demand for efficient vector
database solutions capable of handling high-dimensional semantic search operations. Contemporary infor-
mation retrieval systems face significant challenges in processing complex queries across vast knowledge
repositories while maintaining contextual accuracy and computational efficiency. This research investigates
the optimization potential of vector search implementations in LLMs through comprehensive evaluation using
MongoDB Atlas as the primary vector database platform. Traditional keyword-based retrieval methods fail to
capture semantic relationships and contextual nuances essential for accurate information extraction in modern
AI applications. Vector-based query optimization enables semantic similarity matching, allowing systems to
access contextually relevant data or information even when exact keyword matches are absent. But it signif-
icantly improving response quality and user experience.The study addresses critical performance bottlenecks
in production-scale vector search deployments, where query latency and retrieval accuracy directly impact
system usability. Through systematic comparison of traditional text-embedding-ada-002 against the advanced
text-embedding-3-small model, we demonstrate substantial performance enhancements across multiple eval-
uation metrics. Results establish text-embedding-3-small as superior for semantic search applications, while
GPT-4o-mini demonstrates optimal faithfulness performance (0.9067) for accuracy-critical deployments.
Keywords: Vector Search; Large Language Models; MongoDB Atlas; Semantic Search; Natural Language
Processing; Vector Databases; Embedding Models