Bonsai

bonsai.demo.vide.1.mp4

Bonsai is a multi-agent deep research system, with an emphasis on traceability. Given a query, Bonsai decomposes it into a tree of sub-questions, researches each branch in parallel, and synthesises a final answer — with every intermediate step visible in real time.

Bonsai achieves 65% correctness on the SimpleQA evaluation dataset (n = 20) at a threshold of 0.7 using LLM-as-judge. For comparison, GPT o3 currently sits at 50.5%.

Try out Bonsai live at bonsai.anselmlong.com.

Repository Layout

bonsai/
├── backend/               # FastAPI backend
│   ├── agents/           # Research agents (BranchProcessor, etc.)
│   │   ├── branch_processor.py
│   │   ├── research_graph.py
│   │   └── searcher.py
│   │   └── prompts/      # LangChain prompts
│   │       ├── planner.md
│   │       ├── reflect.md
│   │       └── synthesizer.md
│   ├── models/           # Data models and types
│   │   ├── __init__.py
│   │   └── types.py
│   ├── config.py         # Configuration/settings
│   ├── main.py           # API entry point (FastAPI)
│   ├── Dockerfile
│   └── tests/            # Unit & integration tests
├── frontend/             # Next.js frontend
│   ├── app/              # Next.js app router pages
│   │   ├── page.tsx
│   │   ├── layout.tsx
│   │   └── research/
│   ├── components/       # React components
│   │   ├── AnswerRenderer.tsx
│   │   ├── GraphView.tsx
│   │   ├── NodeDetail.tsx
│   │   ├── QueryInput.tsx
│   │   ├── ResearchTree.tsx
│   │   ├── SourceCard.tsx
│   │   ├── StatusBar.tsx
│   │   └── TreePanel.tsx
│   ├── hooks/            # React hooks
│   │   ├── useResearchStream.ts
│   │   └── useResearchTree.ts
│   ├── lib/              # Shared utilities/types
│   │   └── types.ts
│   ├── public/           # Static assets
│   └── tsconfig.json
├── scripts/              # Utility scripts
│   └── eval.py           # Evaluation script (SimpleQA benchmark)
├── docs/                 # Documentation/iterations
│   ├── ITERATION_1.md
│   └── ITERATION_2.md
├── .env.example          # Environment template
├── .env                  # Environment variables (local)
├── docker-compose.yml    # Docker compose setup
├── pyproject.toml        # Python project config
└── uv.lock               # Python lockfile

Setup

Requirements: Python 3.11+, Node 18+, uv, bun

# 1. Clone and install backend
uv sync

# 2. Set environment variables
cp .env.example .env
# Edit .env with your OPENAI_API_KEY and TAVILY_API_KEY

# 3. Install frontend
cd frontend && bun install && cd ..

Running

Dev Script:

./dev.sh

This development script runs both the backend and frontend simultaneously. If you wish to run them separately, refer to the next 2 commands.

Backend (port 8000):

uv run uvicorn backend.main:app --reload

Frontend (port 3000):

cd frontend && bun dev

Open http://localhost:3000 and enter a research query.

CLI / curl

# Start a research job
curl -s -X POST http://localhost:8000/research \
  -H "Content-Type: application/json" \
  -d '{"query": "What are the long-term economic effects of remote work?"}' | jq .

# Stream events
curl -N http://localhost:8000/research/{job_id}/stream

Tests

# Unit tests (fast, no API keys needed)
uv run pytest backend/tests/ -v -m "not slow"

# Integration tests (requires API keys)
uv run pytest backend/tests/ -v -m slow

Eval

Run the SimpleQA benchmark against a subset of questions:

uv run python scripts/eval.py --n 50 --output results/eval.json

Scores factual accuracy, citation accuracy, completeness, source quality, and conciseness using LLM-as-judge.

Results (SimpleQA, n=20)

Metric	Score
Correct (≥0.7)	13/20 (65%)
Avg latency	30.1s
Avg sources	4.4

Dimension	Score
factual_accuracy	0.70
citation_accuracy	0.88
completeness	0.76
source_quality	0.82
conciseness	0.72

Potential Improvements

This was done as a take home assignment as part of an interview. Given more time (and money), I plan to:

• Use stronger search APIs like Exa
• Could consider searching first, then branching out subqueries from there
• Iterating on the system prompts to get better outputs
• The current graph view is not very smooth, some frontend iterations could fix that too.