Second Wind

Second Wind is an AI workspace for exploring multiple approaches without losing context. Branch from any earlier message, investigate alternatives, and semantically merge the best ideas back together.

Features

• Streaming chat with selectable OpenAI models
• Interactive conversation DAG with checkout and branching
• Semantic three-way merge with common-ancestor and conflict detection
• MongoDB Atlas persistence for conversations, nodes, branches, and merges
• MongoDB Search across saved messages and branch names
• MongoDB Vector Search for relevant ideas from parallel paths
• Multiple conversations with automatic titles, rename, and delete
• File attachments for images, PDFs, text, and code
• SQLite fallback when Atlas is not configured

How Atlas is used

Second Wind stores each message as an individual document in the nodes collection. parent_ids and children_ids preserve the conversation graph, while merge metadata and structured state summaries live beside the message content.

The application uses three collections in the second_wind database:

• conversations — names, timestamps, and the active node
• nodes — messages, graph relationships, branches, summaries, and merges
• attachments — uploaded-file metadata

MongoDB Search powers full-text search. MongoDB Vector Search powers the Second Wind button, which retrieves relevant work from another branch and lets the user select it for a merge.

Run locally

Requirements

• Python 3.10+
• Node.js 20+
• OpenAI API key
• MongoDB Atlas cluster and database user

1. Install the backend

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

Copy .env.example to .env and set the credentials:

OPENAI_API_KEY=your_openai_api_key
MONGODB_URI=mongodb+srv://username:password@cluster.mongodb.net/?retryWrites=true&w=majority
MONGODB_DB=second_wind
MONGODB_SEARCH_INDEX=second_wind_search
MONGODB_VECTOR_INDEX=second_wind_vector
MONGODB_EMBEDDING_MODEL=voyage-4

Never commit .env. URL-encode special characters in the database password.

Start the API:

source .venv/bin/activate
python -m uvicorn server.app:app --reload --host 127.0.0.1 --port 8000

Confirm Atlas is active:

curl http://127.0.0.1:8000/database/status

The response should contain "backend":"mongodb_atlas" and "connected":true. New conversations are then stored directly in Atlas; existing SQLite conversations are not copied automatically.

2. Install the frontend

In another terminal:

cd web
npm install
npm run dev

Open http://127.0.0.1:5173.

3. Create Atlas indexes

In Atlas, open Search & Vector Search for the cluster and use database second_wind, collection nodes:

1. Create second_wind_search using atlas/search-index.json.
2. Create second_wind_vector using atlas/vector-index.json.
3. Wait until each index reports Active.

The vector definition uses Atlas Automated Embedding with Voyage AI. Persistence and normal chat work without the vector index, but semantic Second Wind suggestions require it. See atlas/README.md for the detailed console instructions.

Use the app

1. Create a conversation and select a model.
2. Chat to build the initial path.
3. Click an earlier graph node and continue to create another path.
4. Select two divergent nodes and provide a merge prompt.
5. Enter a new prompt and click Second Wind to retrieve related ideas from other paths.

Stack

React, Vite, FastAPI, OpenAI, MongoDB Atlas, MongoDB Search, and MongoDB Vector Search.