CASE STUDY 02 // HYBRID SEARCH & RAG
Enterprise Smart Search & Document Retrieval
A production AI search system indexing and querying across official documents, videos, website content, FAQs, policies, Scheme Information Documents (SIDs), and published financial literature.
RELEVANCE / ACCURACY
95%+ Precision
RESPONSE TIME
ms to ~1s Latency
SEARCH SCALE
Thousands Daily
RETRIEVAL ENGINE
Milvus + BM25 Fusion
System Highlights & Architecture
Designed to solve the acute limitations of standalone vector search in regulated financial services, where exact acronyms ('ELSS', 'NAV', 'TER') require lexical matching while investor questions require deep semantic intent understanding.
- Hybrid Fusion: Combines dense vector retrieval (Milvus HNSW index) with sparse lexical search (BM25) via Reciprocal Rank Fusion (RRF).
- Cross-Encoder Reranking: High-precision semantic reranker filtering candidate chunks to achieve 95%+ top-3 relevance.
- Table Preservation: Custom document chunking retaining tabular financial data and statutory disclaimers without context loss.
- Real-Time Analytics: Search query logging, intent drift detection, and telemetry tracking popular investor interests.