🎯 What Is Nexus AI?

Nexus AI is a premium meeting intelligence platform that transforms video and audio into structured, actionable knowledge. Drop in a YouTube link or local recording — the system transcribes speech with Groq Whisper, summarizes with Mistral, extracts decisions and action items, and lets you chat with the meeting through a retrieval-augmented pipeline.

Hybrid by design: cloud speech-to-text through Groq keeps transcription fast, while Mistral AI powers summarization, extraction, embeddings, and conversational Q&A.

⚡ Six-Stage Intelligence Pipeline

┌─────────────┐   ┌──────────────┐   ┌─────────────┐   ┌──────────────┐   ┌─────────────┐   ┌──────────┐
│  1. Audio   │ → │ 2. Groq STT   │ → │ 3. Title    │ → │ 4. Summary   │ → │ 5. Extract  │ → │ 6. RAG   │
│  Ingestion  │   │ Transcription│   │ Generation  │   │  (Map-Reduce)│   │  Insights   │   │  Engine  │
└─────────────┘   └──────────────┘   └─────────────┘   └──────────────┘   └─────────────┘   └──────────┘
 yt-dlp / pydub    Groq Whisper API     Mistral LLM        Mistral LLM         Mistral LLM      ChromaDB
 FFmpeg chunking   cloud transcription  8-word title       bullet summary      3 extractors     + Mistral

🧠 Architecture

flowchart TB
    subgraph Input["📥 Input Layer"]
        YT["YouTube URL"]
        LF["Local Video / Audio"]
    end

    subgraph Audio["🎙️ Audio Processing"]
        DL["yt-dlp Download"]
        CV["pydub → 16kHz WAV"]
        CH["10-min Chunks"]
    end

    subgraph STT["🔊 Speech-to-Text — Cloud"]
      GW["Groq Whisper API"]
    end

    subgraph LLM["🤖 Mistral AI + LangChain LCEL"]
        TI["Title Generator"]
        SU["Map-Reduce Summarizer"]
        EX["Action Items · Decisions · Questions"]
    end

    subgraph RAG["💬 RAG Chat"]
        EM["mistral-embed"]
        VS["ChromaDB Vector Store"]
        QA["Context-Aware Q&A"]
    end

    subgraph UI["🖥️ Streamlit Dashboard"]
        ST["Studio"]
        IN["Insights"]
        CHAT["Ask AI"]
    end

    YT --> DL
    LF --> CV
    DL --> CV
    CV --> CH
    CH --> GW
    GW --> TI & SU & EX
    GW --> EM
    EM --> VS
    VS --> QA
    TI & SU & EX & QA --> ST & IN & CHAT

Loading

✨ Capabilities

Module	What You Get
Smart Transcription	Groq Whisper API — fast cloud-based transcription & translation
Executive Summary	Map-reduce summarization across long transcripts
Auto Title	Professional meeting title generated from content
Action Items	Tasks with owner and deadline when mentioned
Key Decisions	Numbered list of decisions made during the meeting
Open Questions	Unresolved topics flagged for follow-up
RAG Chat	Ask natural-language questions grounded in the transcript
Exports	Download transcript, actions, decisions, and questions as .txt
Session History	Timeline of analyses within the current browser session

🖥️ Dashboard

🏠 Home Feature overview & recent activity	⚡ Studio Source input & pipeline launcher	📊 Insights Summary, metrics & extractions	💬 Ask AI RAG-powered meeting chat
🕐 History Session timeline (last 20 runs)	⚙️ Settings Language defaults & session reset	❓ Help Quick start & FAQs	📈 Live Stepper Real-time pipeline progress

🛠️ Tech Stack

Full stack breakdown

Layer	Technology	Role
UI	Streamlit, custom CSS	Multi-page dashboard with glass-morphism theme
Ingestion	yt-dlp, pydub, FFmpeg	YouTube download, format conversion, chunking
STT	Groq Whisper API (whisper-large-v3-turbo)	Cloud transcription & translation
LLM	Mistral AI (mistral-small-2506)	Summaries, titles, extractions, chat
Orchestration	LangChain LCEL	Prompt chains, map-reduce, RAG pipeline
Embeddings	Mistral Embed	Semantic indexing of transcript chunks
Vector DB	ChromaDB	Persistent retrieval store (vector_db/)
Config	python-dotenv	API keys and model settings via .env

📁 Project Structure

Video Agent/
├── app.py                  # Streamlit dashboard entry point
├── main.py                 # CLI pipeline + interactive chat loop
├── requirements.txt
├── .env.example
├── cookies.txt             # Optional — YouTube auth cookies
│
├── core/
│   ├── transcriber.py      # Groq Whisper API transcribe & translate
│   ├── summarizer.py       # Map-reduce summary + title generation
│   ├── extractor.py        # Action items, decisions, questions
│   ├── rag_engine.py       # LCEL RAG chain + ask_question()
│   └── vector_store.py     # ChromaDB build / load / retriever
│
├── services/
│   ├── pipeline.py         # Streamlit pipeline orchestration
│   └── session.py          # Session state, pages, history
│
├── ui/
│   ├── navigation.py       # Sidebar shell & routing
│   ├── components.py       # Reusable UI blocks
│   ├── styles.py           # Global CSS injection
│   └── pages/              # home · studio · insights · chat · history · settings · help
│
├── utils/
│   └── audio_processor.py  # Download, convert, chunk audio
│
└── downloads/              # YouTube audio cache (runtime)

🚀 Quick Start

Prerequisites

• Python 3.10+
• FFmpeg installed and available on PATH
• Mistral API key — console.mistral.ai

Install

git clone https://github.com/SalikAhmad702/Video-Agent.git
cd Video-Agent

# Using uv (recommended)
uv venv
uv pip install -r requirements.txt

# Or standard pip
python -m venv .venv
.venv\Scripts\activate        # Windows
pip install -r requirements.txt

Configure

Copy .env.example to .env and set your keys:

MISTRAL_API_KEY=your_mistral_api_key_here
WHISPER_MODEL=base

Variable	Default	Description
MISTRAL_API_KEY	—	Required for LLM, embeddings, and chat
WHISPER_MODEL	base	Whisper size: tiny, base, small, medium, large-v3
YTDLP_COOKIE_FILE	cookies.txt	Optional path to YouTube cookies for restricted videos

Run

Web dashboard (recommended):

streamlit run app.py

CLI mode:

python main.py

📖 Usage Guide

Web UI

1. Open Studio from the sidebar
2. Paste a YouTube URL or absolute local file path (C:/videos/meeting.mp4)
3. Select your transcription language
4. Click Run Analysis — watch the live pipeline stepper
5. Review output in Insights — summary, transcript, extractions, downloads
6. Switch to Ask AI — chat with suggested prompts or custom questions

CLI

python main.py
# Enter YouTube URL or local file path when prompted
# Chat loop starts automatically after analysis — type 'exit' to quit

Supported Sources

Source	Formats	Notes
YouTube	Any URL	Uses yt-dlp; optional cookies.txt for sign-in videos
Local files	MP4, MP3, WAV, etc.	Converted via pydub → 16 kHz mono WAV

🔬 How It Works

Transcription — Groq Whisper API

Audio is split into 10-minute chunks and sent to the Groq Whisper API (whisper-large-v3-turbo) for fast cloud-based transcription. The translation endpoint is used for non-English audio to produce English output.

Summarization — map-reduce pattern

Long transcripts are split into 3 000-character segments. Each chunk is summarized independently, then a final Mistral pass merges partial summaries into one cohesive bullet-point executive summary.

Extraction — specialized prompts

Three dedicated LangChain chains run against the full transcript:

• Action items — task, owner, deadline
• Key decisions — numbered decision log
• Open questions — unresolved follow-ups

RAG Chat — grounded answers only

The transcript is chunked (500 chars, 50 overlap), embedded with mistral-embed, and stored in ChromaDB. At query time, the top-4 similar chunks are retrieved and passed as context to Mistral — answers stay tied to the meeting content.

⚙️ Configuration Reference

# Required
MISTRAL_API_KEY=sk-...

# Optional — Whisper model size (larger = more accurate, slower)
WHISPER_MODEL=small

# Optional — custom cookie file for YouTube
YTDLP_COOKIE_FILE=C:/path/to/cookies.txt

Whisper model trade-offs:

Model	Speed	Accuracy	RAM
tiny	⚡⚡⚡	★★	~1 GB
base	⚡⚡	★★★	~1 GB
small	⚡	★★★★	~2 GB
medium	🐢	★★★★★	~5 GB

🗺️ Roadmap

• YouTube + local file ingestion
• Groq Whisper API transcription & translation
• Map-reduce summarization with Mistral
• Action items, decisions, and questions extraction
• ChromaDB RAG chat with mistral-embed
• Streamlit dashboard with live pipeline stepper
• Transcript & extraction .txt downloads
• PDF export (reportlab / fpdf2 — deps ready)
• Hindi → English translation layer (deep-translator — dep ready)
• Persistent analysis history across sessions
• Speaker diarization (who said what)
• Docker & Streamlit Cloud deployment guide

🤝 Contributing

Contributions welcome — especially PDF export, speaker diarization, and deployment configs.

git checkout -b feature/your-improvement
git commit -m "feat: describe your change"
git push origin feature/your-improvement

Open a pull request when ready.

📄 License

Free and open source — use, modify, and distribute freely.

• ✅ Personal, academic, and commercial use
• ⚠️ Provided as-is without warranty

---

📧 Let's Connect

Built with obsession by Salik Ahmad 🤖