AI-Powered Search Interface

The Problem

Keyword-based search was returning irrelevant results and unable to understand natural language queries, leading to a 67% search abandonment rate.

The Solution

Built a semantic search interface with streaming LLM-generated summaries, multi-turn conversation support, and real-time relevance feedback that improved the underlying model.

Architecture

Vue 3 frontend consuming Server-Sent Events from a Node.js API that orchestrates OpenAI embeddings and Pinecone vector search.

Key Challenges

• Streaming partial LLM responses to the UI
• Handling search while LLM is generating
• Feeding user relevance signals back to the model