In the ever-expanding world of media consumption, finding new content that matches your taste can be challenging. A new open-source tool called Recommendarr has emerged, generating discussion among tech enthusiasts about the merits of using Large Language Models (LLMs) for personalized media recommendations, as opposed to traditional recommendation algorithms.
AI-Powered Media Recommendations
Recommendarr is a web application that analyzes users' Sonarr, Radarr, Plex, and Jellyfin libraries to generate personalized TV show and movie recommendations. Unlike conventional recommendation systems that rely on collaborative filtering or content-based algorithms, Recommendarr leverages the contextual understanding capabilities of LLMs. The application sends library data to an AI service, which then analyzes viewing patterns and suggests new content based on what it understands about the relationships between different media properties.
The creator of Recommendarr explained their motivation for using LLMs rather than traditional recommendation systems:
I've never found recommendation systems to work very well for me. I've gone through many of them and the reason I decided to start using LLMs was because I was out of options...and after I tried it I ended up much preferring the recommendations given.
This sentiment resonated with several users who expressed frustration with conventional recommendation engines, noting that LLMs can potentially understand nuanced connections between content that might be missed by simpler algorithms.
Knowledge Cutoff Limitations
A significant concern raised in the community discussion is the knowledge cutoff limitation inherent to LLMs. Since these models are trained on data up to a certain point in time, they may not be aware of newer shows or movies released after their training cutoff date. This creates a potential blind spot in recommendations, particularly for recently released content.
The developer acknowledged this limitation, explaining that while models might have some awareness of shows that were upcoming during their training period, recommending very recent releases is likely a weak point of the system. This presents an interesting trade-off: the contextual understanding and natural language capabilities of LLMs versus the ability of traditional recommendation systems to incorporate the latest releases through database updates.
Alternative Approaches and Integrations
Several community members suggested alternative approaches to enhance the recommendation system. One notable suggestion was to use embeddings for clustering rather than relying solely on LLMs. Embeddings could provide a lighter-weight solution that works well with new material by placing media in a multidimensional space where similarity can be measured mathematically rather than through natural language understanding.
Integration with Trakt.tv was another popular suggestion, with users pointing out that this service already integrates with media servers like Emby, Jellyfin, and Plex for many users. The discussion highlighted concerns about how the system would handle extremely large libraries, with some users mentioning collections of over 30,000 movies. The developer noted that such large libraries would likely hit token input limits for LLMs, suggesting a sampling approach as a potential workaround.
Multi-User Household Challenges
A recurring theme in the comments was the challenge of differentiating between different household members' preferences. Many users share their media servers with family members who have vastly different tastes, making unified recommendations less useful. Community members suggested integrating with services like Tautulli and Overseerr to enable per-user recommendations based on individual watching patterns rather than the combined library.
The developer acknowledged this limitation and mentioned that Tautulli integration had been attempted but proved challenging. This highlights one of the key areas where traditional recommendation systems with explicit user profiles might still have an advantage over the current implementation of LLM-based recommendations.
Key Features of Recommendarr
- AI-Powered Recommendations using LLMs
- Sonarr & Radarr Integration for TV shows and movies
- Plex & Jellyfin Integration for watch history analysis
- Flexible AI Support (OpenAI API or compatible alternatives)
- Docker support for easy deployment
- Privacy-focused (credentials stored in browser's local storage)
- Dark/Light Mode UI options
Community-Suggested Improvements
- Per-user recommendations for multi-user households
- Trakt.tv integration for better watch history tracking
- Embedding-based clustering as an alternative to LLMs
- Lidarr integration for music recommendations
- Solutions for handling very large media libraries (30k+ items)
Music Recommendation Potential
Several users expressed interest in extending the concept to music recommendations, suggesting integration with Lidarr (a music collection manager similar to Sonarr and Radarr). One user shared their experience using a script to export their music library for LLM analysis, noting that while imperfect, it provided interesting recommendations. The challenge of LLMs recommending items already in a user's library was mentioned, with a simple but effective solution of explicitly instructing the model not to repeat items from the original list.
As media consumption continues to grow across multiple platforms and formats, tools like Recommendarr represent an interesting exploration of how AI can enhance content discovery. While traditional recommendation systems have been refined over decades, the application of LLMs to this domain offers a fresh approach that may better capture the nuanced relationships between different media properties. The ongoing discussion highlights both the potential and limitations of this approach, suggesting that the ideal recommendation system of the future might combine elements of both traditional algorithms and AI-powered natural language understanding.
Reference: Recommendarr: An AI-driven recommendation system based on Radarr and Sonarr library information