The AI Evolution of Graph Search at Netflix

Netflix evolved its Graph Search platform to support natural language queries using LLMs, translating user input into their internal Graph Search Filter DSL.

• •The previous system required users to construct queries through UI components and a structured DSL, creating friction and inconsistency across applications.
• •The team defined three correctness criteria for generated queries: syntactic (well-formed grammar), semantic (valid fields and controlled vocabulary values, no hallucinations), and pragmatic (captures user intent).
• •Context engineering uses GraphQL schema metadata—field names, types, descriptions, and controlled vocabulary values—as input to the LLM for accurate DSL generation.
• •Controlled vocabularies (finite sets of allowed values defined by domain experts) are extracted and provided as context to prevent the LLM from generating invalid field values.