SamGPT (Sam's Brain) 🧠

Ask one question. Instantly surface the best Sam Parr moments, with timestamped proof.

Why This Exists

Podcast content is high-value and high-volume, but hard to search when you need answers fast.

SamGPT turns long-form conversations into an AI-powered discovery engine:

• Find the exact quote, not just the episode.
• Jump to the exact timestamp, not just a summary.
• Go from research to publishable clips in one workflow.

Example prompts:

• "Find where I found something really funny."
• "Any cool predictions I got right that I can talk about today?"
• "What's a guest idea that actually impressed me?"
• "Find some cool ideas I can repurpose for short form."

How It Works

Under the hood, this is a modern retrieval architecture with clear layers and strong separation of concerns.

flowchart TD
    USER[End User - asks question]
    APP[Web App - captures query and renders results]
    ORCH[AI Orchestration - coordinates the retrieval flow]
    LLM[LLM Layer - intent classification and query expansion]
    VECTORDB[Vector Database - semantic similarity search]
    RANK[Ranking Layer - re-scores and prioritizes best hits]
    CLIP[Async Pipeline - long running clip generation tasks]

    USER --> APP
    APP --> ORCH
    ORCH --> LLM
    ORCH --> VECTORDB
    VECTORDB --> RANK
    RANK --> APP
    APP --> CLIP
    CLIP --> APP

Loading

LLM + Semantic Retrieval Deep Dive

flowchart TD
    Q[User asks a question]
    I[LLM understands intent and goal]
    E[LLM rewrites into better search phrases]
    P[System builds 3 to 15 query phrases]
    M[Each phrase is converted to embeddings]
    V[Vector DB finds similar transcript chunks]
    A[Backend merges and deduplicates matches]
    R[Ranking boosts chunks matched multiple times]
    O[User gets top quotes with timestamps]

    Q --> I
    I --> E
    E --> P
    P --> M
    M --> V
    V --> A
    A --> R
    R --> O

Loading

Async Clip + Download Pipeline (External Backend)

This pipeline is handled by an external backend service (configured via NEXT_PUBLIC_CLIPPING_API_URL), not by this repository.

flowchart TD
    U[User pastes a YouTube link]
    FE[Frontend starts a background clip job]
    API[External backend coordinates the pipeline]
    DL[Worker downloads source media]
    TR[Worker transcribes audio and labels speakers]
    AI[Model layer picks high potential viral moments]
    RENDER[Worker renders final clip segments]
    STORE[Rendered MP4 files are stored]
    READY[Backend marks job complete and returns clip list]
    FILE[Download endpoint streams selected clip file]

    U --> FE
    FE --> API
    API --> DL
    DL --> TR
    TR --> AI
    AI --> RENDER
    RENDER --> STORE
    STORE --> READY
    READY --> API
    API --> FE
    FE --> FILE

Loading

Why It Matters

This project is built for people who care about speed, signal, and shipping:

• Faster insight extraction: ask naturally, get precise moments with context.
• Higher content leverage: turn long episodes into short-form opportunities quickly.
• Production-ready AI stack: orchestration layer, LLM layer, vector retrieval, deterministic ranking, async jobs.
• Real user experience: polished interface, clear loading states, direct links, and clip workflows.

Technology Used

• App framework: Next.js 16, React 19, TypeScript
• UI system: Tailwind CSS v4, Headless UI, Lucide icons
• LLM + embeddings: OpenAI gpt-4o-mini, text-embedding-3-small
• Data platform: Supabase PostgreSQL + pgvector
• Media pipeline integration: external async clipping service (/viral-clips, /clip, /recent-videos)

Get Started

Run it locally:

npm install
npm run dev

Create .env.local:

OPENAI_API_KEY=your_openai_key
SUPABASE_URL=your_supabase_url
SUPABASE_KEY=your_supabase_anon_or_service_key
NEXT_PUBLIC_CLIPPING_API_URL=http://localhost:8000

Then open the app, ask a real question, and explore the moments worth sharing.

---

Built for the MFM community and creator tooling workflows.