Amazon OpenSearch Hybrid Search for RAG Medical Retrieval

The symptom is hybrid-search shaped The reported behavior — missing exact medical terms and acronyms while returning semantically similar but irrelevant documents — points to a dense-vector-only retrieval gap, not an embedding-capacity problem. Amazon OpenSearch Service hybrid search combines vector similarity with keyword matching, so exact tokens such as acronyms and drug names contribute to ranking while semantic similarity still handles paraphrases. That makes A the correct answer, and it does so as a configuration-level change in the existing OpenSearch deployment.

Comments confirm the diagnosis One practitioner described a similar case where vector-only search ranked an employee by name and missed the correct employee ID until the search type was changed to hybrid; another stressed that hybrid search solves both exact-term recall and semantic precision without new services, migrations, or post-processing pipelines. That matches the requirement for low latency and least operational overhead as the collection grows to millions of documents.

Adjacent options add cost without closing the gap B increases embeddings from 384 to 1536 and adds a Lambda post-filter. Larger vectors do not create lexical matching, and the post-filter adds latency and another component to operate. C replaces OpenSearch with Amazon Kendra and applies query expansion; Kendra can be viable when managed enterprise search and built-in connectors are priorities, but migration, reindexing, and application changes are high operational overhead here, and query expansion alone does not directly add exact keyword scoring. D adds a SageMaker AI re-ranker as a second retrieval stage. It may improve precision, but it introduces model hosting, cross-service orchestration, and extra per-query latency, and it cannot recover exact-term matches that the first-stage vector search never returned.

What would flip the decision If the company did not already run OpenSearch, or if a compliance mandate required a managed search service with built-in document connectors, C could become reasonable. If the workload prioritized maximum ranking quality over latency and operational simplicity, D would be more competitive. Under the stated constraints — existing OpenSearch, exact medical terminology, millions of documents, low latency, least operational overhead — hybrid search is the narrowest change that addresses both recall and precision.

Operational consequence of choosing wrong Configuring hybrid search requires tuning lexical and vector weights and monitoring recall and latency, but it avoids a new migration or serving stack. Choosing B or D can make the application slower and more expensive while leaving acronym misses intact; choosing C shifts the operational burden to a migration and a different search product.