The Lenny Growth Assistant

Full-Stack AI Conversational Platform & Artifact Workspace Grounded in Lenny’s Podcast Transcripts

---

Overview

The Lenny Growth Assistant is a full-stack, production-ready AI conversational application built for product managers, founders, and growth leaders. It ingests transcripts from Lenny's Podcast, answers tactical product and growth questions strictly grounded in those transcripts with interactive citation badges, generates structured Ship 30 for 30 atomic essays, and renders interactive sandboxed artifacts (HTML/CSS tools, calculators, frameworks) side-by-side with chat.

flowchart TD
    subgraph Frontend ["Frontend (React + TypeScript + Vite)"]
        UI["Chat Interface & Artifact Viewer"]
        Controls["Model Toggle (Gemini / Ollama) & Sources Drawer"]
    end

    subgraph Backend ["Backend (FastAPI + Async Python)"]
        API["FastAPI REST & SSE Streaming (/api/v1/chat/stream)"]
        Router["Agent Router & Tool Dispatcher"]
        EngineCloud["Google Gemini SDK (3.1 Flash Lite)"]
        EngineLocal["Local Ollama Client (llama3.2 / qwen2.5)"]
        
        subgraph RAG ["Hybrid RAG Engine"]
            Dense["Dense Embeddings (pgvector)"]
            Sparse["BM25 Sparse Search (PostgreSQL FTS)"]
            RRF["Reciprocal Rank Fusion (RRF k=60)"]
            Rerank["Cross-Candidate Reranker"]
        end
    end

    subgraph Database ["PostgreSQL 16 + pgvector"]
        VectorStore["Transcript Vector Database"]
        ChatStore["Sessions, Messages & Artifacts Store"]
    end

    UI <-->|SSE Stream & REST| API
    API --> Router
    Router --> EngineCloud
    Router --> EngineLocal
    Router --> RAG
    Dense <--> VectorStore
    Sparse <--> VectorStore
    API <--> ChatStore

Loading

---

Key Capabilities

1. Hybrid RAG with RRF & Reranking:

   • Dense Semantic Search: Cosine similarity via pgvector embeddings (all-MiniLM-L6-v2).
   • Sparse Keyword Search: PostgreSQL Full-Text Search (tsvector/tsquery) with English dictionary ranking.
   • Reciprocal Rank Fusion (RRF): Merges dense and sparse rankings ($k=60$) for optimal entity & semantic recall.
   • Candidate Reranking: Re-orders top chunks by keyword density and guest-query relevance boost.

2. Speaker-Aware Chunking with Sliding Overlap:

   • Parses YAML frontmatter (guest, title, date, URL) and speaker turns (Lenny: ..., Guest: ...).
   • Slices text into 600-token chunks with 150-token sliding overlap, automatically injecting episode and section headers.

3. Dual Model Engine (Google Gemini & Local Ollama):

   • Cloud Model: Google Gemini 3.1 Flash Lite via official google-genai SDK.
   • Local Model (Mandatory Demo): Local Ollama instance (llama3.2, mistral, qwen2.5) with zero cloud cost.
   • Seamless runtime toggle from the sidebar with automated health-check fallback.

4. Dedicated Ship 30 for 30 Content Skill:

   • Formats grounded knowledge into a ~1,250-word Atomic Essay featuring a strong 1-sentence hook, 1-3-1 narrative cadence, visual skimmability, and tactical playbook.

5. Sandboxed In-App Artifact Viewer:

   • Renders interactive HTML/CSS calculators, dashboards, and Markdown documents in an isolated side panel.
   • Multi-layer security: iframe sandbox="allow-scripts" (strictly omitting allow-same-origin), strict Content-Security-Policy (CSP), and server-side Bleach sanitization.

6. Full Persistence in PostgreSQL:

   • Stores sessions, message threads, latency metrics, citations, and versioned artifacts.

---

Quickstart

Option 1: One-Command Startup with Docker Compose (Recommended)

Make sure Docker and Docker Compose are installed:

# 1. Clone repository and navigate to directory
git clone https://github.com/ShriAmogh/Oogway-Labs-FDE.git
cd Oogway-Labs-FDE

# 2. Copy environment template
cp .env.example .env

# 3. Launch PostgreSQL (with pgvector), FastAPI Backend, and React Frontend
docker compose up --build

• Web Application: http://localhost:3000
• API Docs & Swagger: http://localhost:8000/docs
• Health Endpoint: http://localhost:8000/api/v1/health

---

Option 2: Local Development Mode (Native Python + Node)

Prerequisites

• Python 3.10+
• Node.js 18+ & npm
• PostgreSQL with pgvector running on port 5432 (or run docker compose up -d postgres)
• (Optional) Ollama running locally (ollama run llama3.2)

Step 1: Run PostgreSQL + pgvector

docker compose up -d postgres

Step 2: Start Backend

cd backend
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# Ingest sample transcripts into pgvector
python3 ../scripts/ingest_transcripts.py --sample

# Start FastAPI server
uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

Step 3: Start Frontend

cd frontend
npm install
npm run dev

---

Environment Configuration (.env)

Configure your environment variables in .env:

# Database Configuration (PostgreSQL 16 with pgvector)
DATABASE_URL=postgresql+asyncpg://postgres:postgrespassword@localhost:5432/lenny_growth

# Google AI Studio API Configuration (Cloud Model Provider)
GEMINI_API_KEY=your_gemini_api_key_here
GEMINI_MODEL=gemini-3.1-flash-lite

# Local Ollama Configuration (Local Offline Model Provider)
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=llama3.2

# Default Agent Settings
DEFAULT_PROVIDER=gemini
EMBEDDING_PROVIDER=local
EMBEDDING_MODEL=all-MiniLM-L6-v2

# RAG Search Tuning
RAG_TOP_K_DENSE=20
RAG_TOP_K_SPARSE=20
FINAL_TOP_K=5

---

Running Automated Tests

Run backend unit and integration tests covering chunking with overlap, RRF mathematical fusion, security sanitization, and API endpoints:

cd backend
pytest -v tests/

---

Project Structure

.
├── backend/
│   ├── app/
│   │   ├── agents/          # Gemini SDK Agent, Local Ollama Agent, Router & Tools
│   │   ├── api/v1/          # Chat (SSE), Sessions, Artifacts, Ingestion, Health
│   │   ├── core/            # Config, Async DB with pgvector, JSON Logger, Security
│   │   ├── models/          # SQLAlchemy DB models & Pydantic schemas
│   │   ├── rag/             # Speaker-aware Chunker, Embedder, Hybrid Retriever (RRF)
│   │   └── main.py          # FastAPI app entrypoint with lifespan DB init
│   ├── tests/               # Pytest test suite (RRF, Chunker, Security, Agents)
│   ├── Dockerfile           # Backend Docker container
│   └── requirements.txt
├── frontend/
│   ├── src/
│   │   ├── components/      # Sidebar, Chat, CitationChips, ArtifactViewer, MessageInput
│   │   ├── services/        # Typed API client with resilient SSE parser
│   │   ├── App.tsx          # Resizable split-pane layout & state coordinator
│   │   └── index.css        # Tailwind & Midnight Cyan Mint design tokens
│   ├── Dockerfile           # Frontend multi-stage Nginx build
│   └── package.json
├── docs/
│   ├── PRD.md               # Product Requirements Document & Discovery Brief
│   ├── design.md            # UI/UX Specifications & Iframe Sandbox Security Model
│   ├── architecture.md      # Technical Architecture, DB ERD & RRF Pipeline Flow
│   ├── RAGAS_EVALUATION.md  # Grounding & Faithfulness Benchmark
│   └── CHESKY_EVALUATION_REPORT.md # Comprehensive Evaluation on Chesky Podcast
├── scripts/
│   ├── ingest_transcripts.py # CLI ingestion script for pgvector
│   └── run_local.sh         # One-command local startup script
├── docker-compose.yml       # PostgreSQL (pgvector) + FastAPI + React orchestration
└── .env.example             # Documented environment template

---

Security & Sandbox Isolation

AI-generated HTML artifacts execute in an isolated sandbox:

1. Isolated Iframe: sandbox="allow-scripts" (strictly omits allow-same-origin to prevent access to parent cookies, tokens, and DOM).
2. Content Security Policy: default-src 'none'; style-src 'unsafe-inline'; script-src 'unsafe-inline'; connect-src 'none'; to block unauthorized data exfiltration.
3. Server-Side Sanitization: Python bleach whitelist sanitizes dangerous tags before database persistence.

---

Evaluator Verification Checklist

• Grounded Answers: Ask "What did Brian Chesky say about eliminating traditional PM at Airbnb?" -> Verified response citing Brian Chesky's episode with interactive source pills.
• Out-of-Domain Safety: Ask "What is the recipe for baking sourdough bread?" -> Verified polite refusal acknowledging absence in Lenny's podcast.
• Local Ollama Model: Toggle to Local Ollama -> System queries local llama3.2 via localhost:11434.
• Ship 30 for 30 Skill: Enable Ship 30 for 30 Skill -> Generates ~1,250 word Atomic Essay with hook, 1-3-1 structure, bold highlights, and guest attribution.
• Interactive Artifact Viewer: Ask "Generate an interactive HTML/CSS growth loop calculator" -> Opens side-by-side Artifact Viewer with live executing sandboxed preview and raw code tab.