Niche Horror Recs
In ProgressVibe-based horror recommendations powered by semantic search
The Problem
Most movie recommendation engines are too broad for horror. Genre tags and ratings can tell you if something is horror, but not whether it has the exact mood, tension, weirdness, or dread you are looking for.
The Solution
A recommendation app that uses embeddings and Pinecone vector search to match horror movies by vibe, tone, and feel instead of relying on basic metadata.
Tools
Case Study Snapshot
Search Ranking
70 / 30
Vector similarity blended with niche score to favor deeper cuts over safer mainstream titles.
Theme System
5 Modes
Amber, green, red, cyan, and white themes built around a CRT-style interface.
User Flow
Saved + Random
Search, save to named watchlists, then pull a random pick from mood or list.
Why It Works
The core product decision was to treat horror discovery as a semantic taste problem instead of a genre-filter problem. People are often searching for something emotionally specific, not just "horror."
That shaped both retrieval and ranking. Natural-language search handles detailed prompts directly, while Gemini expands shorter mood phrases into fuller semantic queries before Pinecone retrieval. The ranking blend then biases the results toward cult, obscure, and less obvious titles.
The interface reinforces that niche identity instead of hiding it. The CRT styling, obscurity slider, theme system, and random discovery tools all make the app feel like a purpose-built recommendation machine rather than a generic movie database.
Search System
- Natural-language search for prompts like slow-burn atmospheric grief film.
- Pinecone integrated llama-text-embed-v2 model for query embeddings.
- Gemini expands short mood prompts into richer semantic search queries.
- Results ranked by 70% vector similarity and 30% niche score.
Film Card Experience
- Poster art pulled from TMDb.
- AI-generated one-sentence explanation for why each film matches the query.
- Niche tier badge such as Deep Cut, Cult Pick, or Hidden Gem.
- Expandable details with director, streaming providers, and external links.
Retention + Personalization
- Named watchlists with add, remove, and random-pick behavior.
- Random film discovery from a mood query or a saved list.
- Email and password auth with a seven-day JWT session.
- Logged-in behavior uses watch history as reranking context.
Interface Direction
The UI leans into a terminal and CRT-inspired visual system with scanlines, strong color themes, and a responsive drawer-based mobile layout. That aesthetic choice is not just style. It helps the product feel specific, memorable, and aligned with the kind of niche exploration the recommendation engine is built for.
Architecture / Flow