Conversational product search for a B2B marketplace

Problem

The catalog covered machines described by dozens of technical parameters in a complex taxonomy. Customers struggled to find a machine matching their needs through traditional filters and keyword search.

Solution

We built a conversational agent replacing the classic search bar on the site. The system recognizes user intent, asks follow-up questions about missing requirements (machine specifications), and then searches the product database combining structured search (SQL over parameters) with vector search (semantic matching), falling back to external sources (web search) when the internal catalog doesn’t have enough matches. The agent can also take actions on the user’s behalf — adding a product to favorites or sending an inquiry to the supplier.

The architecture is built from several specialized sub-agents (intent router, search, site API interactions, response preparation) implemented in LangGraph, with full LLM call tracing in LangSmith (OpenAI and Anthropic models). Machine data is collected by scraping manufacturer websites, cleaned and normalized, with embeddings generated for every machine.

We also built an evaluation framework for the response quality of each sub-agent — based on a golden dataset and manual review by domain experts in a dedicated interface, where every answer could be marked correct/incorrect along with a comment.