🪄🎓 VideoRAC: Retrieval-Adaptive Chunking for Lecture Video RAG

🏛️ Official CSICC 2025 Implementation

"Adaptive Chunking for VideoRAG Pipelines with a Newly Gathered Bilingual Educational Dataset"

(Presented at the 30th International Computer Society of Iran Computer Conference — CSICC 2025)

---

📊 Project Pipeline

---

📖 Overview

We present Video-RAC, an adaptive chunking methodology for lecture videos within Retrieval-Augmented Generation (RAG) pipelines. Using CLIP embeddings and SSIM to detect coherent slide transitions, plus entropy-based keyframe selection, we construct multimodal chunks that align audio transcripts and visual frames.

Alongside the method, we release EduViQA, a slide-centric, bilingual (Persian/English) lecture dataset containing 20 videos from 5 professors across STEM and education topics. Each lecture is paired with 50 synthetic QA items and categorized by duration (40% mid-length, ~20–40 minutes) to support controlled RAG benchmarking.

This repository is the official implementation of the CSICC 2025 paper by Hemmat et al.

Hemmat, A., Vadaei, K., Shirian, M., Heydari, M.H., Fatemi, A. “Adaptive Chunking for VideoRAG Pipelines with a Newly Gathered Bilingual Educational Dataset.” Proceedings of the 30th International Computer Society of Iran Computer Conference (CSICC 2025), University of Isfahan.

---

🧠 Research Background

This framework underpins the EduViQA bilingual dataset, designed for evaluating lecture-based RAG systems in both Persian and English. The dataset and code form a unified ecosystem for multimodal question generation and retrieval evaluation.

Key Contributions:

• 🎥 Adaptive Hybrid Chunking — Combines CLIP cosine similarity with SSIM-based visual comparison.
• 🧮 Entropy-Based Keyframe Selection — Extracts high-information frames for retrieval.
• 🗣️ Transcript–Frame Alignment — Synchronizes ASR transcripts with visual semantics.
• 🔍 Multimodal Retrieval — Integrates visual and textual embeddings for RAG.
• 🧠 Benchmark Dataset — 20 bilingual educational videos with 50 QA pairs each.

---

📊 Dataset

EduViQA: Bilingual Educational Video QA Dataset

Dataset composition highlighting topic distribution and lecture duration proportions.

Dataset Statistics

Metric	Value
Total Videos	20 (10 Persian, 10 English)
Professors	5
Duration Focus	40% mid-length (20–40 minutes)
QA Pairs per Video	50 synthetic QA pairs
Format	JSON annotations

Topics Covered

• Computer Architecture
• Data Structures
• System Dynamics and Control
• Teaching Skills
• Descriptive Research
• Regions in Human Geography
• Differentiated Instruction
• Business

The dataset also captures slide transitions and keyframes extracted via CLIP+SSIM chunking, enabling multimodal retrieval experiments with aligned visuals and transcripts.

📥 Access Dataset: Hugging Face - EduViQA

---

⚙️ Installation

pip install VideoRAC

---

🚀 Usage Example

1️⃣ Hybrid Chunking

from VideoRAC.Modules import HybridChunker

chunker = HybridChunker(
    clip_model='openai/clip-vit-base-patch32',
    alpha=0.6,
    threshold_embedding=0.85,
    threshold_ssim: float=0.8,
    interval: int=1,
)
chunks, timestamps, duration = chunker.chunk("lecture.mp4")
chunker.evaluate()

2️⃣ Q&A Generation

from VideoRAC.Modules import VideoQAGenerator

def my_llm_fn(messages):
    from openai import OpenAI
    client = OpenAI()
    response = client.chat.completions.create(model="gpt-4o", messages=messages)
    return response.choices[0].message.content

urls = ["https://www.youtube.com/watch?v=2uYu8nMR5O4"]
qa = VideoQAGenerator(video_urls=urls, llm_fn=my_llm_fn)
qa.process_videos()

---

📈 Results Summary (CSICC 2025)

Method	AR	CR	F	Notes
VideoRAC (CLIP+SSIM)	0.87	0.82	0.91	Best performance overall
CLIP-only	0.80	0.75	0.83	Weaker temporal segmentation
Simple Slicing	0.72	0.67	0.76	Time-based only

Evaluated using RAGAS metrics: Answer Relevance (AR), Context Relevance (CR), and Faithfulness (F).

---

🧾 License

Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).

You may share and adapt this work with attribution. Please cite our paper when using VideoRAC or EduViQA:

@INPROCEEDINGS{10967455,
  author={Hemmat, Arshia and Vadaei, Kianoosh and Shirian, Melika and Heydari, Mohammad Hassan and Fatemi, Afsaneh},
  booktitle={2025 29th International Computer Conference, Computer Society of Iran (CSICC)}, 
  title={Adaptive Chunking for VideoRAG Pipelines with a Newly Gathered Bilingual Educational Dataset}, 
  year={2025},
  volume={},
  number={},
  pages={1-7},
  keywords={Measurement;Visualization;Large language models;Pipelines;Retrieval augmented generation;Education;Question answering (information retrieval);Multilingual;Standards;Context modeling;Video QA;Datasets Preparation;Academic Question Answering;Multilingual},
  doi={10.1109/CSICC65765.2025.10967455}}

---

👥 Authors

University of Isfahan — Department of Computer Engineering

• Kianoosh Vadaei — kia.vadaei@gmail.com
• Melika Shirian — mel.shirian@gmail.com
• Arshia Hemmat — amirarshia.hemmat@kellogg.ox.ac.uk
• Mohammad Hassan Heydari — heidary0081@gmail.com
• Afsaneh Fatemi — a.fatemi@eng.ui.ac.ir

---

Star History

---

⭐ Official CSICC 2025 Implementation — Give it a star if you use it in your research! ⭐ Made with ❤️ at University of Isfahan