AI-Powered Search at Scale

At Frame.io (Adobe), we built AI-powered search to help creative professionals find the exact shot or moment across millions of assets. This required two AI integrations: an LLM for natural language queries and vector embeddings for semantic visual search. Our LLM component uses Meta’s Llama 4 via AWS Bedrock, translating queries like “videos of people on beaches from last week” into OpenSearch DSL. For visual search, we generate vector embeddings using Adobe’s SearchCut API (CLIP-based models), enabling similarity searches without metadata. The challenge? Video embedding takes 30+ seconds per asset. Our Oban infrastructure—already split across four instances due to database contention—couldn’t handle this load without impacting critical jobs. Our solution: a multi-stage Broadway pipeline consuming from SQS. A router pipeline handles authorization and versioning, fanning out to dedicated embedding and transcription workers. This provides independent scaling, traffic buffering, and complete isolation from our job system. You’ll learn: practical LLM integration patterns, when to choose Broadway over Oban, and how to build observable AI pipelines in Elixir.