https://github.com/antoniakras/semantic-video-search

GPU-optimized semantic search on video transcripts, with benchmarking of FAISS, Pinecone, and PostgreSQL vector databases. Deployed via Docker on FORTH’s GPU infrastructure.
https://github.com/antoniakras/semantic-video-search

bert-embeddings bert-fine-tuning cuda dokcer embedding-models embeddings-word2vec faiss-vector-database gpu-computing huggingface-transformers nlp-machine-learning pgvector pineconedb postgresql python pytorch retrieval-augmented-generation similarity-search vector-database whisper-ai

Last synced: about 1 month ago
JSON representation

GPU-optimized semantic search on video transcripts, with benchmarking of FAISS, Pinecone, and PostgreSQL vector databases. Deployed via Docker on FORTH’s GPU infrastructure.

• Host: GitHub
• URL: https://github.com/antoniakras/semantic-video-search
• Owner: antoniakras
• Created: 2025-09-20T19:24:37.000Z (9 months ago)
• Default Branch: main
• Last Pushed: 2025-09-30T10:17:08.000Z (9 months ago)
• Last Synced: 2025-09-30T12:18:21.794Z (9 months ago)
• Topics: bert-embeddings, bert-fine-tuning, cuda, dokcer, embedding-models, embeddings-word2vec, faiss-vector-database, gpu-computing, huggingface-transformers, nlp-machine-learning, pgvector, pineconedb, postgresql, python, pytorch, retrieval-augmented-generation, similarity-search, vector-database, whisper-ai
• Homepage:
• Size: 11.7 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# 🎓 Bachelor Thesis: GPU-Optimized Semantic Search on Video Transcripts

## 📌 Overview

This repository documents my **Bachelor of Science in Computer Science thesis project** at **FORTH (Foundation for Research and Technology – Hellas)**.  

The project implemented an **end-to-end Retrieval-Augmented Generation (RAG) pipeline**, enabling **semantic search over video transcripts** with results enriched by **timestamps and video identifiers**. 

✨ Key features include:  

- **RAG Architecture**: Combined video transcript retrieval with embedding-based semantic search.  

- **End-to-End Workflow**: Built a **Django REST Framework** API for embedding generation, vector storage, and similarity search.  

- **Semantic Video Transcript Search**: Extracted audio from videos, transcribed with **OpenAI Whisper**, and embedded with **BERT**.  

- **GPU Optimization**: Leveraged **PyTorch + FAISS GPU** for high-performance similarity search and efficient memory use.

- **Vector Database Benchmarking**: Conducted a comparative study of **FAISS, Pinecone, and PostgreSQL**  

- **Embedding Optimization**  

- **Deployment**: Containerized the system with **Docker** for reproducibility and portability.  

- **Environment**: Developed in **Python on Linux Ubuntu**, running exclusively on **FORTH’s GPU infrastructure via VPN**.

⚠️ **Note**: The source code is not publicly available due to infrastructure and legislation restrictions. This repository serves as a **portfolio showcase**. Access for review under supervision is possible by request.

---

## 🛠️ Tech Stack

- **Languages**: Python  

- **Frameworks/Models**: PyTorch, Hugging Face Transformers (BERT), OpenAI Whisper, Django REST Framework  

- **Vector Databases**: FAISS (GPU), Pinecone, PostgreSQL (pgvector)  

- **Benchmarking**: FAISS GPU benchmarking tools + custom evaluation scripts  

- **Deployment**: Docker, Linux Ubuntu  

- **Other Tools**: NLTK (text preprocessing), MoviePy (audio extraction)

## 🚀 Features

- End-to-end **RAG pipeline** for semantic retrieval.  

- Transcribes video to text with **Whisper**.  

- Encodes transcripts into embeddings with **BERT**.  

- Optimized embeddings with **normalization and quantization**.  

- Stores and queries embeddings in **vector databases**.  

- Retrieves **most semantically similar transcript segments** with their **timeframes and source video IDs**.  

- Fully **GPU-accelerated** for optimized performance.  

- Exposed as a **REST API** for seamless integration.  

---

## 📊 Vector Database Benchmarks & Embedding optimization 

**Embedding Optimization**:  

  - Applied **normalization (L2, quantization, etc.)** for space and time efficiency.  

  - Matched **embedding normalization strategies** with appropriate similarity search methods (inner product, cosine similarity, Euclidean distance). 

**Vector Databse Benchmarks**

I tested and compared FAISS, Pinecone, and PostgreSQL on:  

- **Accuracy and integrity** of top-k retrieved results  

- **Query speed / latency** and API response time (time complexity)  

- **Storage efficiency** under different embedding optimization strategies (space complexity)

**Each benchmark combined**:  

- Embedding normalization choice.  

- Appropriate similarity search metric.

---

## 📸 Screenshots 

**Workflow**



---

## About This Project

- **Type**: Bachelor Thesis Project  

- **Institution**: University of Crete, Computer Science Department  

- **Research Infrastructure**: The Foundation for Research and Technology - Hellas (FORTH)