Skip to content
Conference

A Retrieval-Augmented Generation Framework for Query-Driven Video Content Grounding

Oct 2026 · International Conference on Innovative Mechanisms for Industry Applications · pp. 926-930 · 0 citations · 23 references

Abstract

Video Question Answering (Video-QA) requires answers grounded in visual events, temporal context, and spoken or written evidence. Conventional large language models can generate plausible but unsupported responses when long videos are compressed into incomplete prompts. This paper presents a lightweight Video-RAG framework that retrieves relevant video segments before generation. Videos are divided into fixed 8 -second chunks; CLIP embeddings are extracted from representative frames, BERT/T5 embeddings are extracted from ASR transcripts or subtitles, and both streams are indexed with FAISS. At inference, a hybrid visual-textual similarity score retrieves timestamped evidence that is provided to a multimodal LLM. The framework is validated on standard Video-QA settings, including MSR-VTT-QA, TVQA, and ActivityNet-QA, using retrieval quality, answer accuracy, memory, and runtime metrics. Results show that dual visual-textual retrieval improves grounding over text-only, visual-only, Standard LLM, CLIP, and Video-LLaMA baselines while remaining practical on commodity GPU and Apple Silicon hardware. The work demonstrates that explicit retrieval is an effective and deployable mechanism for factual, explainable Video-QA.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.