A RAG-based music recommendation engine (CORRD)
Music recommendation traditionally rests on the collaborative signal: what people similar to you have listened to. That signal is missing exactly where it would matter most — the rarely played album, the new release, the user who just arrived — and on a catalog combining local library and streaming, the long tail is the interesting part. On top of that, a recommendation that cannot say why it was made convinces few of the people who choose music deliberately.
Analysis
Survey of the descriptive content available across the catalog — album, artist and track metadata, genres, relationships between artists, editorial content — and assessment of its coverage, because an engine reasoning over content is worth only as much as the content it can reach. Definition, together with the product team, of what makes a recommendation good in this context: which is not the same as making it correct.
Solution
A recommendation engine that indexes the catalog's descriptive content and retrieves by semantic similarity from what the user is playing or from a natural-language request, building the suggestion on the retrieved content. Because retrieval works on descriptions rather than on co-listening, an album can be recommended with no listening history at all: precisely the point where collaborative filtering has nothing to say.
Architecture
Semantic indexing of catalog metadata and editorial content, vector search for similarity, generation constrained to items actually present in the catalog and available to the user — recommending a record that does not exist or cannot be played is worse than recommending nothing — and traceability of the content behind each suggestion.
Implementation
Started on a subset of the catalog, compared against the recommendation logic already in place and evaluated with real users: in music, the gap between metrics computed offline and the recommendations people actually accept is wider than elsewhere.
Duration
From kickoff to first production release: about three months.
Measured results
- 85% of recommendations judged relevant by real users in the evaluated sample
- 65% catalog coverage reached with the new logic, compared with the previous recommendation logic
- Recommendations possible even for albums with no listening history, where the collaborative signal does not exist
- Every suggestion traces back to the content that produced it, so it can be explained rather than left opaque
- Suggestions constrained to the catalog actually available to the user, not to plausible but absent titles
- Natural-language requests handled without depending on a rigid genre taxonomy
ROI: Value is read in the catalog depth users actually reach, in reduced dependence on accumulated listening history, and in how well recommendations hold up on the newer or less-travelled part of the catalog.
Scope and measurement method
- Scope: the subset of the catalog indexed at the start, progressively extended to the rest.
- Evaluation set: recommendations judged by real people, not offline metrics alone.
- What is monitored: relevance of the suggestion, diversity, and catalog coverage. They have to be read together because they move in opposite directions: an engine maximizing relevance ends up recommending the same things over and over.
- Recommendation quality is not accuracy: there is no correct answer to compare against. That is why evaluation needs people, and an evaluation set alone cannot tell you whether the engine works.
- Measured relevance: 85% of recommendations judged relevant by users in the evaluated sample. Measured coverage: 65% of the catalog reached compared with the previous recommendation logic. Catalog diversity remains tracked separately and has no published figure yet: high relevance achieved by always recommending the same titles would not be progress.
Client: Project delivered for Volumio: it powers the recommendation engine. volumio.com/