GloZ's Case Study on Building a Natural Language Resume Search System Based on Amazon OpenSearch Service — Part 2: Hybrid Search and Natural Language Query Transformation

When searching for translators, complex natural language conditions such as "native English/Spanish speaker with dubbing experience" were entered, but existing simple keyword matching or rule-based parsing had limitations in accuracy. Additionally, metadata filtering and vector search were separated, making it difficult to consistently handle complex conditions such as AND/OR/NOT.

To solve this, the following solution was built.

• Hybrid Search Optimization: Used Amazon OpenSearch's normalization-processor to normalize BM25 and Vector scores that had different scales. Grid Search was used to derive the optimal weight combination, improving search performance.
• Natural Language → DSL Conversion: Used Claude Haiku 4.5 on Amazon Bedrock via Function Calling. The LLM only extracts search parameters while the backend assembles the DSL, ensuring consistency of results.
• RAG Pattern Application: To improve conversion accuracy, past conversion cases were stored in OpenSearch, and similar cases were injected as Few-shot examples.