Hybrid Semantic Vector Search

Problem Keyword-based search engines (like BM25) fail when users query using synonyms, natural language questions, or when they express implicit intents.

Solution Designed a hybrid search engine combining vector embeddings (\`paraphrase-MiniLM\`) and Elasticsearch BM25 search. Integrated an LLM-based query intent classifier that reads queries and dynamically adjusts the weights of vector vs. keyword scores.

Challenges * **Vocabulary Mismatch**: Users searching for "automobile coverage" instead of "car insurance". * **Solution**: Dense embeddings easily captured semantic similarity. * **Latency & Payload**: Running transformers for every query can choke servers. * **Solution**: Quantized embeddings and optimized Elasticsearch indexes.

Results * Instantaneous semantic matching. * Relevance scores (NDCG@10) increased significantly. * High intent routing accuracy.