Knowledge Graph Enhancement with LLMs for Content Understanding

netflix · Media & Entertainment · year n/a · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Entertainment content knowledge graph

Netflix needed to enrich and maintain a large entertainment knowledge graph by matching entities at massive scale and extracting relationships not explicitly present in source data.

Constraints:

• Entity matching spans billions of pairs.
• The pipeline must run across distributed infrastructure.
• New LLM-based relationship inference must preserve production stability.
• Unstructured content contains semantic relationships not available as explicit fields.

Architecture

Netflix uses Metaflow to orchestrate large-scale entity matching and is exploring RAG plus in-context learning to infer entity types and relationships from unstructured content.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  entertainment-source-data["Entertainment Source Data"]
  metaflow-entity-matching["Metaflow Entity Matching"]
  apache-arrow-loading["Apache Arrow Loading"]
  llm-relationship-inference["LLM Relationship Inference"]
  knowledge-graph["Knowledge Graph"]
  entertainment-source-data --> metaflow-entity-matching
  metaflow-entity-matching --> apache-arrow-loading
  metaflow-entity-matching --> knowledge-graph
  apache-arrow-loading --> metaflow-entity-matching
  llm-relationship-inference --> knowledge-graph

Components

Data flow

Entertainment data flows into Metaflow-orchestrated distributed jobs for entity matching, fast Apache Arrow loading accelerates pair processing, and newer LLM/RAG pipelines infer semantic types and relationships from unstructured content.

Why this architecture

Metaflow provides distributed orchestration and stability for massive matching, while LLM/RAG inference targets semantic relationships unavailable in structured source fields.

Models used: LLMs · RAG pipelines · Metaflow · Apache Arrow

Scale

• Entity matching runs across billions of pairs.
  “billions of entity matching pairs”summary, stated
• Apache Arrow data loading produced 10x speedups.
  “achieving 10x speedups through fast data loading with Apache Arrow.”summary, stated

Alternatives considered (roads not taken)

Explicit source-data relationships only

Rejected because: Some semantic relationships are not explicitly available in the source data.

Ad hoc distributed jobs without Metaflow isolation

Rejected because: Production stability required namespacing and branching features.

Patterns

rag-retrieval (retrieval)

Netflix uses RAG pipelines to support LLM relationship extraction from unstructured content.

durable-execution (workflow)

Metaflow orchestrates massive distributed entity-matching jobs with production namespacing and branching.

Tradeoffs

Use Metaflow orchestration

Chose Namespaced and branched distributed pipelines over Unmanaged distributed matching scripts.

Rationale: The production graph required stability while running enormous matching jobs.

Add LLM inference to graph enrichment

Chose RAG and in-context learning over Only explicit structured relationships.

Rationale: LLMs can infer semantic relationships not directly present in the source data.

Gotchas (what bit them)

Knowledge graph enrichment can require relationships absent from source schemas.

Root cause: Unstructured content encodes semantic relationships implicitly.

Resolution: Explore LLM/RAG in-context inference to extract entity types and relationships.

Large-scale entity matching is bottlenecked by data loading.

Root cause: Billions of pair comparisons stress distributed IO.

Resolution: Use Apache Arrow fast loading for a 10x speedup.

Evaluation

The summary reports system speedup and production-stability practices but does not give relationship-extraction accuracy metrics.

Metrics: Entity-matching speedup · Distributed pipeline stability

Outcomes

• Apache Arrow loading produced 10x speedups.
• Metaflow orchestrates billions of entity-matching pairs.
• LLM/RAG approaches are being explored for semantic relationship inference.

The knowledge graph workflow operates over billions of entity-matching pairs.