Semantic Search in Healthcare: Inside Halodoc’s RAG Retrieval Architecture

Halodoc, Indonesia's leading healthcare app, shares how they built a semantic retrieval layer on top of their existing keyword-based search to power RAG-driven AI systems like a prescription assistance platform and Healthcare Q&A assistant. Using OpenSearch with HNSW-based k-NN vector search and 768-dimensional embeddings from Hugging Face sentence-transformers, they created an ingest pipeline that automatically generates embeddings at indexing time. The architecture complements rather than replaces keyword search, forming a hybrid retrieval system. Key lessons include the importance of offline embedding generation for scale, hybrid retrieval improving results, and retrieval quality directly impacting hallucination rates. The system has validated ~21K prescriptions and driven ~32K bookings from Q&A assistant users.