Aipower Inc.
Hybrid-Retrieval Sales Chatbot
Dense + sparse retrieval for e-commerce guided selling and troubleshooting.
Jun 2025 – Oct 2025
Hybrid retrieval that stopped an e-commerce chatbot from fumbling SKUs
Dense-only semantic search failed exactly where e-commerce answers matter most: exact terms like SKUs and model numbers, where embedding similarity blurs the tokens that distinguish products. Adding a sparse keyword path alongside the dense one fixed the failure at its source and cut latency versus the dense-only baseline. The requirements came from discovery with sales and marketing, who knew which queries were losing customers.
- Dense + sparse
- hybrid retrieval
- ↓ latency
- vs dense-only baseline
- ↑ precision
- on SKU / model-number queries
- RAG
- Hybrid Retrieval
- LLMs
- Vector Search
- Python
The problem
Dense-only semantic retrieval failed exactly where e-commerce answers matter most: exact-term queries like SKUs and model numbers, where embedding similarity blurs the very tokens that distinguish one product from another.
What I built
- 01Architected hybrid retrieval combining dense semantic search with sparse keyword search, so exact identifiers stayed exact while natural-language intent still resolved.
- 02Deployed into AI sales chatbots covering both guided selling and product troubleshooting.
- 03Led technical discovery directly with marketing and sales, translating their customer requirements into AI solution designs and architecture decisions.
Outcome
Improved answer precision on exact-term queries and cut response latency versus the dense-only baseline. Delivered the contracted product ahead of the department wind-down.