🎬👻 CineMantic – Search by Vibe or Memory, Not Title

Ever wanted to find a movie with a certain vibe?

A setting. A feeling. Like:

• Surreal horror set in remote Scottish mountains

Or maybe you remember fragments of something you watched years ago, like:

• Antarctic alien that imitates humans
• Girl crawling out of a TV

…but not the name.

CineMantic lets you search for movies the way humans actually remember them —
by vibes, scenes, and fragments of plot, not exact titles.

It’s not just search. It’s a vibe check for your watchlist.

---

🌐 Live Demo:

https://cinemantic.onrender.com/

⚠️ Initial load may take a moment due to cold starts on free-tier hosting and the demo may occasionally be unavailable due to API rate limits or service tier restrictions.

---

• Search example

• Search example

---

💡 Why I Built This?

I love horror movies.

But I kept running into the same problem:

• Sometimes I’d be in the mood for something specific — atmospheric, isolating, psychological.
• Other times, I’d remember bits of a movie from years ago… but not the title.

Forums and Google searches were hit-or-miss and time-consuming.

Traditional search expects keywords.
Human memory doesn’t work like that.

So I built 🎬👻 CineMantic — a small experiment in semantic movie discovery.

---

✨ What It Does?

• 🧠 Understands plot fragments and conceptual descriptions from user input
• ⚡ Uses vector similarity search to retrieve the closest matches
• 🎬 Displays posters, TMDB ratings, runtime, and release dates
• 🌙 Clean, dark-themed interface built with Streamlit

Instead of typing exact names, you describe what you feel or remember.

---

🧩 How It Works (Under the Hood)?

The system follows a retrieval-first architecture inspired by modern RAG pipelines, separating query understanding, vector search, and metadata enrichment.

1️⃣ Query Normalization (LLM-Assisted)

User input is normalized using an LLM to bridge the gap between fuzzy human descriptions and structured TMDB-style movie overviews.

This improves alignment before embedding and helps semantic retrieval perform more accurately.

---

2️⃣ Semantic Embeddings

Both user queries and movie overviews are transformed into dense vector representations using a transformer-based embedding model.

These vectors capture conceptual similarity beyond simple keyword overlap.

---

3️⃣ High-Dimensional Vector Search (FAISS)

All movie overview embeddings are indexed using FAISS for efficient nearest-neighbor search.

When you search:

• Your query is embedded into the same vector space
• FAISS retrieves the top-k nearest neighbors
• Results are ranked by similarity score

---

4️⃣ Metadata Enrichment Layer

Matched vector IDs are joined with a lightweight SQLite database to retrieve posters, ratings, runtime, and release data.

---

🔍 Example Query Flow

Input:
psychological thriller with shocking plot twists

Process:

• Query → normalized via LLM
• Query → embedding vector
• FAISS → top k nearest overviews
• SQLite → retrieve metadata

Output:
Ranked movies with similarity scores and full context.

---

🎃 Dataset & Scope

This prototype currently indexes:

Top 500 highest-rated horror films on TMDB

While the demo focuses on horror, the retrieval architecture is genre-agnostic and can scale to larger datasets.

---

🤖 Models Used

CineMantic uses Mistral models for both:

• LLM-assisted query normalization

• Text embeddings generation

---

📌 Disclaimer

This product uses the TMDB API but is not endorsed or certified by TMDB.

🚀 Running Locally

git clone https://github.com/ABFirest0rm/cinemantic.git
cd cinemantic
pip install -r requirements.txt
streamlit run app.py

⚠️ Before running locally, create a .env file:

MISTRAL_API_KEY=your_key_here

Use your own API key. Model names or API calls may need small adjustments depending on what’s available to you.

📄 License

This project is licensed under the MIT License.