I just pushed v22 of my project : a local AI companion for Radarr, that goes beyond generic genre or TMDb lists.
This isn’t “yet another recommender”. It’s your personal taste explorer that actually gets the vibe you want in natural language and builds recommendations starting from your existing library.
Key highlights from a real recent run:
- Command:
--mood "dystopian films like Idiocracy, Gattaca or In Time" - Output: Metropolis (1927), V for Vendetta, Children of Men, Brazil (1985), Minority Report, Dark City, Equilibrium, Upgrade, The Road… → oppressive/surveillance/inequality/societal critique atmosphere, not just “dark sci-fi”.
How it works :
- Starts by sampling random movies from your Radarr collection (or uses your mood/like/saga input).
- Asks a local Ollama LLM (e.g. mistral-small:22b) for 25 thematic suggestions based on atmosphere/vibe.
- Validates each via OMDb (IMDb rating, genres, plot, director, cast…).
- Scores intelligently: IMDb rating + genre match + director/actor bonus + plot embedding similarity (cosine on Ollama embeddings).
- Adds the top ones directly to Radarr (with confirmation: all / one-by-one / no).
- Persistent blacklist to avoid repeats.
Different modes :
--mood "dark psychological thrillers with unreliable narrators", any vibe you describe--like "Parasite" --mood "mind-bending class warfare"(or just--like "Whiplash")--saga(auto-detects incomplete sagas in your library and suggests missing entries) or--saga "Star Wars"--director "Kubrick"/--actor "De Niro"/--cast "Pacino De Niro"(movies where they co-star)--analyze→ full library audit + gaps (e.g. “You’re missing Kurosawa classics and French New Wave”)--watchlist→ import from Letterboxd/IMDb--auto→ perfect for daily cron / Task Scheduler (wake up to 10 fresh additions)
Standout features:
- 100% local + privacy-first (Ollama + free OMDb API only)
- No cloud AI, no tracking
- colored console output, logs, stats, HTML/CSV exports
- Synopsis preview before adding
- Configurable quality profile, min IMDb, availability filters
- Works on Windows, Linux, Mac
GitHub (clean single-file Python script + detailed README):
https://github.com/nikodindon/radarr-movie-recommender
If you’re tired of generic Discover lists, Netflix-style randomness, or manual hunting give it a spin. The vibe/mood mode + auto saga completion really change how you expand your collection.
Let me know what you think, any weird mood examples you’d like to test, or features you’d want added!
Since no one is leaving critical comments that might explain all the downvotes, I’m going to assume they’re reflexively anti-AI, which frankly, is a position that I’m sympathetic to.
But one of the benign useful things I already use AI for, is giving it criterias for shows and asking it to generate lists.
So I think your project is pretty neat and well within the scope of actually useful things that AI models, especially local ones, can provide the users.
No LLM use is benign. The effects on the environment, the internet, and society are real, and that cannot be ignored.
You can make the argument that in some cases it is justified, e.g.: for scientific research.
Saw it was already commented about CO2, so I thought I’d counter-point your environment claim regarding water usage (since that is something I’ve seen a lot of too).
The ISSA had a call to action due to the AI water use “crisis”: https://www.issa.com/industry-news/ai-data-center-water-consumption-is-creating-an-unprecedented-crisis-in-the-united-states/
68 billion gallons of water by 2028! That’s a lot…right? Well, what I found is that this is somewhat of a bad faith argument. 68 billion gallons annually is a lot for one town, but those are numbers from a national level and it isn’t compared to usage from anything else. So, lets look at US agriculture (that’s something that’s tracked very well by the USDA): https://www.nass.usda.gov/Publications/Highlights/2024/Census22_HL_Irrigation_4.pdf
That’s 26.4 trillion gallons of water annually. So, AI datacenter represents 0.26% of agriculture consumption. If AI datacenter consumption is a crisis, why is agriculture consumption not a crisis? You could argue that agriculture produces “something useful”, but usefulness doesn’t factor into the scarcity of a resource. So, either its not a crisis, or you are cherry picking something that has no meaningful outcome to solving the problem.
chill, this is extracting text embeddings from a local model, not generating feature-length films
that’s like saying “no jet use is benign” meant for comparing a private jet to a jet-ski
the generative aspect is not even used here
The effects on the environment
Didn’t down vote you. I hear this line of complaint in conjunction with AI, especially if the person saying it is anti-AI. Without even calculating in AI, some 25 million metric tons of CO2 emissions annually from streaming and content consumption. Computers, smartphones, and tablets can emit around 200 million metric tons CO2 per year in electrical consumption. Take data centers for instance. If they are powered by fossil fuels, this can add about 100 million metric tons of CO2 emissions. Infrastructure contributes around 50 million metric tons of CO2 per year.
Now…who wants to turn off their servers and computers? Volunteers? While it is true that AI does contribute, we’re already pumping out some significant CO2 without it. Until we start switching to renewable energy globally, this will continue to climb with or without AI. It seems tho, that we will have to deplete the fossil fuel supply globally before renewables become the de facto standard.
The more local inference, the better. Nice work!
Thanks for the discussion ! I Added more details on the various modes. Apparently I sparked some controversy , I understand some people don’t like AI. I find the tool pretty cool to use, feel free to test it and report anything you want ^^
Anti-AI evangelism is at its peak rn.
that’s pretty cool, this is just the wrong crowd, don’t worry about the downvotes
How does this compare to an ML approach?
are you training or just using an LLM for this?
There’s no training, the LLM embeddings are used to compare the plots via a cosine similarity, then a simple weighted score with other data sources is used to rank the candidates. There’s no training, evaluation, or ground-truth, it’s just a simple tool to start using.
Y’all need LLMs to tell you what to watch, now?
