Movie Recommendation System using Oracle Vector Search

Overview

This project implements a semantic movie recommendation system using Oracle Database with native vector support.

Instead of traditional filtering (genre, rating), the system uses:

• text embeddings
• cosine similarity
• vector search (ANN)

to recommend movies based on semantic similarity.

---

Solution Architecture

TMDB API (external data)
│
▼
Ingestion Script (Deno + TypeScript)
│
▼
Oracle Database (Docker container)
│
├─ Store movie data (relational + JSON)
├─ Generate embeddings → VECTOR(384) (ONNX model)
├─ Create vector index (ANN)
└─ Similarity search (cosine distance)
│
▼
Web Server (Hono)
│
▼
Client (Browser UI)

---

Steps

1. Load the embedding model into Oracle

The all-MiniLM-L6-v2 model (ONNX format) is copied into the Docker container and loaded into Oracle. This allows embeddings to be generated directly inside the database.

---

2. Load movie data

Movie data is fetched from the TMDB API using the ingestion script and inserted into the database.

---

3. Prepare text input

title + genres + overview

---

4. Generate vectors

UPDATE movies
SET embedding = VECTOR_EMBEDDING(mini_lm_model USING text_input);

Each movie is transformed into a 384-dimensional embedding vector.

---

5. Create vector index

CREATE VECTOR INDEX movies_emb_idx
ON movies (embedding)
DISTANCE COSINE;

This enables efficient similarity search.

---

6. Query similarity

SELECT *
FROM movies
ORDER BY VECTOR_DISTANCE(
embedding,
(SELECT embedding FROM movies WHERE id = :id),
COSINE
)
FETCH FIRST 10 ROWS ONLY;

Returns the most similar movies.

---

Score Interpretation

Score	Meaning
0.00	Identical movie
< 0.05	Almost identical
< 0.15	Highly similar
< 0.30	Moderately similar
< 0.50	Weak similarity
≥ 0.50	Different movies

Notes

• Score = cosine distance
• Lower = more similar

---

Data Model

CREATE TABLE movies (
id NUMBER PRIMARY KEY,
title VARCHAR2(255),
overview CLOB,
release_date VARCHAR2(20),
poster_path VARCHAR2(255),
genres JSON,
vote_average FLOAT,
embedding VECTOR(384)
);

---

Technology Stack

Layer	Technology	Purpose
Database	Oracle 23c	storage + vector search
ML Model	ONNX (MiniLM)	embeddings
Backend	Deno + Hono	server
Language	TypeScript	logic
Data Source	TMDB API	movie data
Containerization	Docker	runtime
Orchestration	Docker Compose	service management

---

Features

• Semantic movie recommendations
• Vector embeddings stored in DB
• Cosine similarity search
• ANN vector index
• Full pipeline (ingestion → embedding → query)