Natural language search lets users ask questions in everyday language and receive results from Netflix's Graph Search platform without writing a Graph Search Filter DSL statement.

Problem Statement & Motivation

Users interact with dozens of UI components across Content and Business Products, each requiring manual construction of DSL filters. This creates friction and inconsistent experiences.

• Multiple bespoke query builders increase learning overhead.
• Hundreds of index fields make UI forms cumbersome.
• SMEs must translate domain knowledge into technical syntax.
• Inconsistent DSL support leads to errors.

LLM‑Powered Text‑to‑DSL Engine

The core engine uses a large language model to convert free‑form questions into syntactically valid Graph Search Filter DSL statements.

• Prompt design balances instruction clarity with token efficiency.
• Model selection prioritizes low latency and high accuracy.
• Output is constrained by a JSON schema to enforce grammar.
• Supports ambiguous phrasing through iterative clarification.

Context Engineering & Schema Extraction

Accurate DSL generation requires the LLM to understand field names, types, and controlled vocabularies derived from the GraphQL schema.

• Automated schema parser extracts field metadata.
• Controlled vocabularies are injected as enumerated value lists.
• Metadata includes description, type, and permitted values.
• Context payload is cached for fast reuse.

Validation & Post‑Processing Pipeline

Generated statements undergo multiple checks to ensure they meet syntactic, semantic, and pragmatic criteria.

• Syntax validator parses the DSL with the official grammar.
• Semantic checker cross‑references field types and allowed values.
• Pragmatic layer runs a mock query against a sandbox index to verify intent alignment.
• Feedback loop surfaces confidence scores to the UI.

Deployment & Operational Considerations

The solution runs as a self‑managed microservice integrated with existing Netflix applications.

• Containerized deployment using Kubernetes for autoscaling.
• Observability via metrics, logs, and trace IDs.
• Rate limiting protects LLM usage costs.
• Rollout includes A/B testing against legacy DSL builders.

For reference on building resilient services, see the real‑time orchestration framework and the scalable data platform guides.