The AI Evolution of Graph Search at Netflix

Alex Hutter, Bartosz Balukiewicz27 Jan 2026

Problem

Netflix's internal content and business applications filter data tables through a custom filter language (DSL) built on the Graph Search platform. Each application collects input in its own way, so users must learn every UI. Some domains have hundreds of filterable fields, and a subject-matter expert who knows the question still has to translate it into form fields and facets. The friction slows them down or blocks them.

Example from the article: "I want to see all movies from the 90s about robots from the US" has to become a valid filter over an index with hundreds of fields.

Solution in brief

An LLM generates the filter statement, but not from the whole schema. Vector search over fields and over controlled-vocabulary values picks a small context, the LLM writes the DSL, and deterministic checks (parser plus metadata lookup) reject hallucinations. The generated filter is shown back as editable UI chips and facets, and users can pin entities with @mentions.

Questions this case answers
  • How to translate an ambiguous natural-language question into a valid structured filter when the schema is too large to fit in the LLM context
  • How to detect and prevent hallucinated fields and values in LLM-generated queries without adding latency
  • How to let non-technical users verify and correct a machine-generated query before trusting its results
  • How to let non-technical users filter enterprise data in plain language instead of a query language or a large form?

Tools and methods

  • Field RAG — retrieves the few index fields a question refers to.
  • Controlled-vocabulary RAG — retrieves matching allowed values and the fields they belong to.
  • AST parser for the DSL — checks syntax and field and value names.
  • Chips and facets in the UI — show the generated filter for review and editing.
  • @mentions — let the user fix an entity, skipping inference.
Continue reading — it's free

Google or a one-time email code. No password.