Advancing semantic search for millions of Rovo users

This post details advancements in Rovo's semantic search capabilities, focusing on the evolution of embedding models (MiniLM, BGE-large, EmbeddingGemma-300m) and the rationale behind their selection based on quality, latency, and cost. It explains the hybrid retrieval approach combining traditional signals with neural semantic understanding for improved relevance. The post also highlights task-aware search tailoring results to user context (Jira, Confluence) and discusses future investments in deeper work context modeling, proactive assistive behavior, tenant-specific embedding fine-tuning, and the integration of AI agents. A significant technical contribution is the collaboration with NVIDIA, leveraging their accelerated computing and AI software stack (Llama-Nemotron-Embed-1B-V2, NeMo Data Designer, NeMo Automodel) to achieve a 26%-40% uplift in retrieval quality through fine-tuning. The post also mentions the use of NVIDIA NeMoClaw and OpenShell runtime for agentic AI.