Building a RAG-Enhanced Recommendation System

First working version of Swanson integrated into whatisonthe.tv.

Key design decisions

SSE over WebSockets: Simpler for unidirectional streaming with native browser support and automatic reconnection.

Natural language + pattern parsing: Streaming natural text with structured extraction at the end rather than forcing JSON responses.

Taste profile building: Aggregates genre counts, content types, and recent titles from the 50 most recent check-ins to inject context into prompts.

The feedback loop

Recommendations incorporate previous feedback within a session. Ask for comedies, get suggestions, mark some as “not interested”, ask again - Swanson remembers and adjusts.

Full write-up

Read the complete build post on swm.cc for architecture details, code samples, and future considerations.