RAG-Powered Video Intelligence

Why did Video A
outperform Video B?

Paste two social media video URLs. VideoRAG extracts transcripts, computes engagement metrics, embeds content into a vector store, and gives you streaming AI answers with cited sources.

Vector-embedded transcripts
Streaming GPT-4o responses
Source-cited answers
2.8s
Avg. Analysis Time
ada-3-small
Embedding Model
512 tok
Vector Chunks

Compare Two Videos

Paste one YouTube link and one Instagram link below

VS
A

Video A — YouTube

Paste a YouTube video or Shorts URL

YouTube

Supports: youtube.com/watch, youtu.be, youtube.com/shorts

B

Video B — Instagram

Paste an Instagram Reel or video URL

Instagram

Supports: instagram.com/reel, instagram.com/p (public only)

Processing takes ~3 seconds · Transcripts embedded with text-embedding-3-small

Processing Pipeline

What happens when you submit

Fetch Metadatayt-dlp

Pull views, likes, comments, creator info, hashtags

Extract Transcriptyoutube-transcript-api

Download captions or generate via Whisper

Chunk & Embedtext-embedding-3-small

Split transcript into 512-token chunks, embed each

Index in Vector DBChromaDB / Qdrant

Store chunks with video_id tags (A or B)

RAG ReadyLangChain + GPT-4o

LangChain retriever initialized with memory

Architecture: LangChain → ChromaDB → GPT-4o streaming → memory buffer chain

Questions you can ask after analysis

RAG + Citations

How it works

6 core capabilities

Streaming Responses

GPT-4o streams tokens in real-time via Server-Sent Events. No waiting for full response.

Source Citations

Every claim cites the exact video and chunk. Click any citation pill to see the source text.

Conversation Memory

LangChain ConversationBufferMemory maintains context across all turns in your session.

LangChain Orchestration

RetrievalQA chain with MMR retrieval. Tagged chunks ensure A/B context separation.

Vector Store

Chunks embedded with text-embedding-3-small, stored in ChromaDB with video_id metadata filter.

Engagement Analytics

Computed engagement rate = (likes + comments) / views × 100, displayed with delta comparison.