Spotify's data assistant Vedder represents a different approach to AI-powered data analysis. Instead of giving an LLM raw access to database schemas and hoping for the best, the team built a context layer that encodes domain expertise[reference:62]. This context layer is the secret sauce that makes the assistant reliable.
Why Context Matters
Raw database schemas are terrible inputs for LLMs. They're full of cryptic table names, obscure column names, and complex relationships that no model can interpret correctly. Query history is better, but it's messy and full of one-off queries that worked for one person one time. Neither approach scales to thousands of users asking thousands of questions.[reference:63]
Spotify's solution is to have domain experts curate the context[reference:64]. These experts know what the data means. They know which tables are authoritative. They know which queries produce correct results. They encode this knowledge into a context layer that the AI assistant can use to answer questions reliably.[reference:65]
The Cluster Model
The context layer is organized around a "cluster model"[reference:66]. Different teams curate the context for their domains. The recommendation team curates data about recommendations. The playback team curates data about playback. This distributed ownership ensures that the context stays accurate and up to date.
The Results
The results speak for themselves[reference:67]. Over 2,100 Spotifiers have used Vedder in more than 13,000 conversations[reference:68]. Users ask questions in simple English and get reliable data within seconds[reference:69]. The assistant has democratized access to insights across thousands of fast-moving teams[reference:70].
The Future of Enterprise AI
Spotify's context layer approach points to the future of enterprise AI. The models will keep getting better, but they'll still need high-quality context to be useful[reference:71]. The companies that win will be the ones that figure out how to encode domain expertise into context layers that AI assistants can use.
Vedder shows that this is possible at scale[reference:72]. It's not just about building a better model. It's about building a better context.