Semantic–Lexical Fusion: Improving Retrieval Accuracy for AI-Driven Knowledge Systems

Abstract

Large language models increasingly rely on external knowledge retrieval to generate accurate, context-aware responses. Dense vector representations enable semantic similarity matching but struggle with exact term identification, structured metadata constraints, and domain-specific identifiers common in enterprise environments. This framework integrates dense embedding-based retrieval with token-level inverted indexing to address these limitations. The architecture employs dual indexing structures, a unified query decomposition layer, and a weighted scoring mechanism that combines cosine similarity with BM25 relevance signals. Field-level filtering and access control mechanisms ensure compliance with organizational constraints while maintaining semantic generalization.

Performance characterization across production-scale deployments demonstrates that hybrid architectures achieve 60-150ms query latency with 1.0-2.0GB memory footprint per million documents, while pure semantic approaches require 768MB-1.5GB and pure lexical systems consume 200-500MB. Index construction analysis reveals vector encoding demands 15-50 hours for million-document collections compared to minutes for inverted indices, though batch processing strategies mitigate operational impact. Sensitivity analysis across weighting parameters identifies optimal semantic-lexical balance at α = 0.6, achieving peak F1 score of 0.847 and demonstrating 15-30% accuracy improvements over single-paradigm methods.

Evaluation across open-domain corpora (MS MARCO, Natural Questions) and enterprise document collections demonstrates improved precision and recall, particularly for queries containing product identifiers, user entities, and structured filters. The system supports retrieval-augmented generation pipelines, conversational interfaces, and knowledge-grounded chatbots where both conceptual relevance and deterministic matching are essential. Results indicate that fusing complementary retrieval paradigms provides robust performance across diverse query types while maintaining interpretability and production-grade reliability for AI-driven applications. Cost analysis for enterprise deployments reveals 50-80% infrastructure premium for hybrid systems compared to single-approach implementations, justified by reduced hallucinations, improved user satisfaction, and decreased support escalations in operational AI applications.