Encoding Your Domain Expert: The Context Layer Behind Spotify's Data Assistant

2026-06-10 · Source: Spotify Engineering · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, medium

Summary

Spotify developed "Vedder," an AI data assistant, to address the overwhelming demand for data insights from its 70,000+ datasets, which process 1.4 trillion data points daily. Traditional LLM approaches failed due to limited context windows and schemas not conveying critical business logic. Vedder, actively utilized since August 2025 by over 2,100 Spotifiers across 13,000+ conversations, operates on a "cluster model" where domain experts curate "clusters" of data. Each cluster comprises relevant datasets with full schemas, vetted question-and-SQL example "pairs," and additional business "docs." This human-curated context is crucial for trustworthiness; a trial showed experts accepted only 12.5% of automatically generated question-SQL pairs from query history, highlighting the noise in raw data. Clusters are continuously monitored via health scores, prompting experts to update context as data evolves.

Key takeaway

For AI Architects or Data Scientists building internal data assistants, relying solely on raw schemas or query logs for LLM context is insufficient and untrustworthy. You should empower domain experts to curate and own specific data "clusters" with vetted examples and business context. This approach ensures accuracy and scalability, transforming experts from answering one-off questions to shaping a reliable knowledge layer that serves thousands. Continuously monitor context health to prevent degradation and maintain trust.

Key insights

Human-curated context, not raw schemas, is essential for trustworthy AI data assistants at scale.

Principles

• Data experts must own context curation.
• Raw query history contains significant noise.
• Trustworthy AI requires human judgment.

Method

Spotify's data agent uses a ReAct loop, selecting context, writing SQL, running queries, and returning answers with sources.

In practice

• Implement a "cluster model" for domain-specific context.
• Monitor cluster health scores to maintain context validity.
• Integrate user feedback to refine knowledge bases.

Topics

• AI Data Assistant
• Context Curation
• LLM Applications
• Data Governance
• SQL Generation
• ReAct Framework

Best for: AI Product Manager, Product Manager, CTO, AI Engineer, Data Scientist, AI Architect

Related on AIssential

• Are LLMs safe? — LLM training data curation, especially via quality filters, embeds implicit biases that shape model behavior and capabilities.
• Data Products: The Essential Context for Enterprise AI — Enterprise AI agent success hinges on robust, architected context layers, not just advanced models.
• I Don’t Use Spotify. So I Built My Own AI DJ. — Building a personalized AI music curator requires overcoming API limitations and enriching diverse, often obscure, music metadata.
• Give Your AI Unlimited Updated Context — An LLM-maintained personal wiki creates a persistent, compounding knowledge artifact, eliminating repetitive context setup.
• Your AI Doesn’t Know Anything. And That’s Not the Model’s Fault. — AI hallucinations stem from infrastructure limitations, not model intelligence, necessitating robust knowledge layers like graphs.
• How OpenAI Built Its Data Agent — The reliability of AI agents at scale stems from strong data foundations and carefully engineered context, not complex agent architectures.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Spotify Engineering.