Steam Similarity Recommender adds taste vectors

This is a smart take on game discovery: it treats recommendations as taste matching, not just genre matching, which is exactly where Steam’s tag system gets too coarse.

• –Mining reviews for niche cues gives the recommender a better shot at surfacing why a game feels right, not just what it is
• –The explainability angle is the real product value here; “you liked the city vibe, jazz, and social loop” is more useful than “similar users also played this”
• –The main risk is data quality: review text is noisy, subjective, and can drift badly if the pipeline is stale or overfits rare phrasing
• –The 3-day ETL and “data may be 3 months old” limitation suggest freshness is the bottleneck, not the ranking model
• –Chroma plus PostgreSQL is a practical stack for this kind of semantic retrieval product, but the long-term moat will be the tag taxonomy and curation quality, not the infra