Lenny's Roundtable

A RAG-powered app where podcast guests answer your questions based on what they actually said on Lenny's Podcast, then discuss with each other.

🚀 Hosted Demo

No installation required. The live demo uses a pre-built vector database hosted on HuggingFace — users skip the 50-minute local ingest step entirely.

---

What You Get

• 300+ transcripts from Lenny's Podcast included (Thanks to https://www.lennysnewsletter.com/ - go subscribe, now)
• 292 unique speakers including product leaders from Stripe, Netflix, Airbnb, and more
• Multi-round discussions where experts build on each other's perspectives

Quick Start

Step 1: Clone and install

git clone https://github.com/manimohans/lennysroundtable.git
cd lennysroundtable
uv sync

Step 2: Configure your LLM

Copy the example config:

cp .env.example .env

Edit .env with your provider:

Option A: Ollama (Local, Free) - Tested ✓

1. Install Ollama
2. Pull models:

   ollama pull embeddinggemma   # For embeddings
   ollama pull gemma3:4b        # For generation (or any model you prefer)

3. Set in .env:

   LLM_BASE_URL=http://localhost:11434/v1
   LLM_MODEL=gemma3:4b
   LLM_API_KEY=not-needed
   

Option B: LM Studio (Local, Free) - Tested ✓

1. Install LM Studio
2. Download a model and start the local server
3. You still need Ollama for embeddings:

   ollama pull embeddinggemma

4. Set in .env:

   LLM_BASE_URL=http://localhost:1234/v1
   LLM_MODEL=your-loaded-model-name
   LLM_API_KEY=not-needed
   

Option C: OpenAI (Cloud, No Ollama Required) - Tested ✓

1. Get an API key from OpenAI
2. Set in .env:

   LLM_BASE_URL=https://api.openai.com/v1
   LLM_MODEL=gpt-4o-mini
   LLM_API_KEY=sk-your-api-key-here
   
   EMBEDDING_PROVIDER=openai
   OPENAI_API_KEY=sk-your-api-key-here
   

   With EMBEDDING_PROVIDER=openai, Ollama is not required — embeddings use text-embedding-3-small instead.

Step 3: Generate embeddings (one-time setup)

Why isn't this included? The vector database (chroma_db/) is ~600MB, which exceeds GitHub's file size limits. You need to generate it locally once. This takes about 50 minutes on CPU.

If using Ollama embeddings (default):

ollama pull embeddinggemma    # If you haven't already
uv run python -m roundtable.ingest

If using OpenAI embeddings (EMBEDDING_PROVIDER=openai in .env):

uv run python -m roundtable.ingest

No Ollama needed — embeddings are generated via the OpenAI API (text-embedding-3-small).

You'll see progress like:

Found 301 transcript files
[301/301] Parsing: Zoelle Egner.txt

Created 17482 parent documents
Unique speakers: 292

Storing parent documents...
Creating child chunks for vector indexing...
Embedding and indexing child chunks (this may take a while)...
Generating embeddings: 100%|████████| 2048/2048 [04:08<00:00]
...

Ingestion complete!
Parent chunks: 17482
Child chunks: 22435

Step 4: Run the app

uv run streamlit run roundtable/app.py

Open http://localhost:8501

---

How to Use the App

1. Enter your details

• Your Name: How speakers will address you (default: "PM")
• Your Question: Ask anything about product, startups, leadership, etc.

2. Adjust settings (sidebar)

Setting	Description
Discussion Rounds	How many rounds of back-and-forth (1-5)
Number of Experts	How many speakers to include (3-7)
Response Length	1=brief, 5=detailed

3. Generate and read

Click Generate Discussion and watch:

1. Initial Thoughts - Each expert shares their perspective
2. Round 2+ - Experts respond to each other, agree/disagree, add nuance

4. Save your discussion

Click Save as Markdown to download the full conversation.

---

Sync Latest Transcripts (Optional)

Want to update with the newest podcast episodes?

Step 1: Get Dropbox access

1. Go to https://www.dropbox.com/developers/apps
2. Create an app with "scoped access" and "full Dropbox"
3. In Permissions tab, enable: sharing.read, files.content.read
4. Generate an access token
5. Add to .env:

   DROPBOX_ACCESS_TOKEN=your_token_here
   

Step 2: Sync transcripts

uv run python sync_transcripts.py

Step 3: Re-embed

uv run python -m roundtable.ingest --reset

---

Deploying to Streamlit Cloud

Host your own live demo in three steps.

Step 1: Build and upload the vector database

Run the ingest pipeline locally first (see Quick Start above), then upload the resulting chroma_db/ to HuggingFace:

pip install huggingface_hub
huggingface-cli login          # once — creates ~/.cache/huggingface/token

python scripts/upload_chroma.py --repo-id YOUR_HF_USERNAME/lennysroundtable-chroma

This creates a public HuggingFace dataset containing chroma_db.tar.gz (~600 MB).

Step 2: Fork & connect to Streamlit Cloud

1. Fork this repo on GitHub.
2. Go to share.streamlit.io → New app.
3. Point it to roundtable/app.py in your fork.

Step 3: Add secrets

In Settings → Secrets on Streamlit Cloud, paste:

LLM_BASE_URL = "https://api.openai.com/v1"
LLM_MODEL    = "gpt-4o-mini"
LLM_API_KEY  = "sk-..."

EMBEDDING_PROVIDER = "openai"
OPENAI_API_KEY     = "sk-..."

HF_REPO_ID = "YOUR_HF_USERNAME/lennysroundtable-chroma"

On first boot the app downloads the pre-built ChromaDB (~1–2 min) and caches it for the lifetime of the server instance. Subsequent users see no delay.

See .streamlit/secrets.toml.example for the full template.

---

Architecture

RAG Pipeline

User Question
     │
     ▼
┌─────────────────────────────────────┐
│  Embed query (Ollama or OpenAI)     │
│  Search child chunks (512 chars)    │
│  Return parent chunks (2048 chars)  │
│  Rank speakers by relevance         │
└─────────────────────────────────────┘
     │
     ▼
┌─────────────────────────────────────┐
│  Generate responses (your LLM)      │
│  Stream to UI                       │
│  Build on previous responses        │
└─────────────────────────────────────┘

Parent-Child Retrieval

We use a two-tier chunking strategy:

• Child chunks (512 chars): Small pieces for precise semantic matching
• Parent chunks (2048 chars): Larger coherent chunks returned to the LLM

This gives you accurate retrieval AND rich context.

Speaker Scoring

Speakers are ranked by relevance to your question:

1. Cosine similarity: Each child chunk gets a similarity score (0-1, higher = more relevant)
2. Aggregation: Sum all similarity scores for a speaker's matching chunks
3. Normalization: Divide by √(num_chunks) to balance speakers with many vs. few appearances

score = sum(similarities) / sqrt(num_chunks)

Typical scores range from ~0.5 (weakly relevant) to ~2.0 (highly relevant). The top 5 speakers are selected for the discussion.

Embedding Model

Two providers are supported, configured via EMBEDDING_PROVIDER in .env:

Provider	Default model	Requires
ollama (default)	embeddinggemma	Ollama running locally
openai	text-embedding-3-small	OPENAI_API_KEY

Override the model with EMBEDDING_MODEL. Important: the embedding model must stay consistent between ingestion and queries — if you re-ingest with a different model, delete chroma_db/ first.

---

Project Structure

lennysroundtable/
├── roundtable/
│   ├── app.py           # Streamlit UI
│   ├── generator.py     # LLM response generation
│   ├── retriever.py     # Semantic search & ranking
│   ├── ingest.py        # Embedding pipeline
│   └── parser.py        # Transcript parsing
├── transcripts/         # 300+ podcast transcripts
├── chroma_db/           # Generated locally (not in repo)
├── .env.example         # Configuration template
└── sync_transcripts.py  # Fetch new episodes

---

Configuration Reference

LLM Settings (.env)

Variable	Description	Example
LLM_BASE_URL	API endpoint	http://localhost:11434/v1
LLM_MODEL	Model name	gemma3:4b
LLM_API_KEY	API key (if needed)	not-needed or sk-...

Embedding Settings (.env)

Variable	Default	Description
EMBEDDING_PROVIDER	ollama	Embedding backend: ollama or openai
EMBEDDING_MODEL	embeddinggemma (ollama) / text-embedding-3-small (openai)	Embedding model name
OPENAI_API_KEY	—	Required when EMBEDDING_PROVIDER=openai
OLLAMA_BASE_URL	http://localhost:11434	Ollama host (ollama provider only)

Chunk Settings (roundtable/ingest.py)

Variable	Default	Description
CHILD_CHUNK_SIZE	512	Smaller = more precise matching
PARENT_CHUNK_SIZE	2048	Larger = more context for LLM

---

Troubleshooting

"No relevant speakers found"

• You need to run the embedding step first: uv run python -m roundtable.ingest
• If using Ollama: make sure it's running with embeddinggemma pulled

"Connection refused" errors

• Make sure your LLM provider is running (Ollama/LM Studio)
• Check the LLM_BASE_URL in .env

Embedding step fails (Ollama)

• Ensure Ollama is running: ollama serve
• Pull the embedding model: ollama pull embeddinggemma
• Or switch to OpenAI embeddings: set EMBEDDING_PROVIDER=openai and OPENAI_API_KEY in .env

Embedding step fails (OpenAI)

• Check that OPENAI_API_KEY is set in .env
• Verify the key is valid at https://platform.openai.com/api-keys

Slow generation

• Local models depend on your hardware
• Try a smaller model like gemma3:1b or use OpenAI

---

Tech Stack

• RAG Framework: LlamaIndex
• Vector Store: ChromaDB
• Embeddings: embeddinggemma via Ollama (default) or text-embedding-3-small via OpenAI
• LLM: Any OpenAI-compatible API
• UI: Streamlit
• Package Manager: uv

---

License

MIT