databricks-agent-toolkit 0.3.0

Unified toolkit for building production agents on Databricks - zero-frontend with official UI templates (Streamlit, Gradio, Dash)

Databricks Agent Toolkit

Build production agents on Databricks in minutes, not days.

Generate production-ready agent scaffolds that follow official Databricks patterns and work with any of the 6 official UI templates.

pip install databricks-agent-toolkit

# Generate a chatbot (use the short alias!)
dbat generate chatbot my-bot
cd my-bot
python start_server.py

# 🎉 Done! Agent running at http://localhost:8000

---

Why This Toolkit?

Philosophy: "On Top Of, Not Instead Of"

We don't create custom frameworks. We generate production-ready agent backends that follow:

• ✅ OpenAI API standard (industry-wide compatibility)
• ✅ Official Databricks patterns (FastAPI + OpenAPI)
• ✅ MLflow best practices (auto-tracing, experiments)
• ✅ Databricks integrations (Model Serving, Lakebase, Unity Catalog)

Your agents work with:

• All 6 official Databricks UI templates
• Any OpenAI-compatible UI framework
• Custom frontends (React, Streamlit, Gradio, etc.)

---

Quick Start

1. Install

pip install databricks-agent-toolkit

2. Generate an Agent

# Use the short alias (much easier to type!)
dbat generate chatbot my-bot

# Or the full command:
# databricks-agent-toolkit generate chatbot my-bot

Available aliases:

• dbat → Databricks Agent Toolkit ✨ (recommended)
• dat → Even shorter!
• databricks-agent-toolkit → Full command

3. What You Get

my-bot/
├── agent.py              # Your agent logic (OpenAI API compatible)
├── start_server.py       # FastAPI server with auto-docs
├── chatbot.py           # CLI interface for testing
├── config.yaml          # Easy configuration
├── requirements.txt     # All dependencies
├── databricks.yml       # Deploy to Databricks Apps
├── app.yaml            # App configuration
└── README.md           # Setup instructions

4. Run Locally

cd my-bot
pip install -r requirements.txt

# Start the server
python start_server.py
# 🚀 Backend: http://localhost:8000
# 📚 API Docs: http://localhost:8000/docs
# 💚 Health: http://localhost:8000/health

# Or test in CLI
python chatbot.py

5. Deploy to Databricks

# One command deployment
databricks bundle deploy

# Your app is live at: https://<workspace>/apps/my-bot

---

What's Included

L1: Chatbot (Simple Conversational AI)

dbat generate chatbot my-bot

Features:

• 💬 OpenAI API compatible via MLflow AgentServer (/invocations)
• 🔄 Streaming with Server-Sent Events (SSE)
• 📊 MLflow auto-tracing (requires mlflow>=3.6.0)
• 🎛️ Configuration-driven (no code changes to switch models)
• 🚀 One-command deploy to Databricks Apps
• 📝 Built-in web UI (or use official templates)
• 📖 Auto-generated OpenAPI docs

Perfect for:

• Quick prototypes and demos
• Simple Q&A bots
• Customer support assistants
• Internal tools

---

L2: Assistant (With Memory + RAG)

dbat generate assistant my-assistant --enable-rag

Everything in L1, plus:

• 💾 Conversation memory (Lakebase/PostgreSQL)
• 🧠 RAG with pgvector or Databricks Vector Search
• 📁 Auto-index documents from Unity Catalog Volumes
• 🔐 Session management
• 👥 Multi-user support

Perfect for:

• Knowledge base assistants
• Document Q&A
• Support bots with history
• Enterprise applications

---

Configuration

Edit config.yaml to customize your agent:

# config.yaml
model:
  endpoint: databricks-meta-llama-3-1-70b-instruct  # Any Databricks Model Serving endpoint
  temperature: 0.7
  max_tokens: 500
  streaming: true
  token_delay_ms: 50  # Streaming speed (lower = faster)
  system_prompt: "You are a helpful AI assistant."

mlflow:
  experiment: /Shared/my-bot
  auto_trace: true  # Automatic tracing of all LLM calls

# L2 only: Memory configuration
memory:
  enabled: true
  backend: lakebase
  host: ${LAKEBASE_HOST}
  database: ${LAKEBASE_DATABASE}

# L2 only: RAG configuration
rag:
  enabled: true
  source: /Volumes/main/default/docs  # Unity Catalog Volume
  backend: pgvector  # or vector_search
  embedding_model: databricks-bge-large-en

No code changes needed! Just edit config and redeploy.

---

Official UI Templates

Your agent backend is 100% compatible with all official Databricks UI templates:

Framework	Best For	Template
Streamlit	Quick prototypes, data apps	streamlit-chatbot-app
Gradio	ML demos, simple interfaces	gradio-chatbot-app
Plotly Dash	Data dashboards with chat	dash-chatbot-app
Shiny	Statistical apps, R users	shiny-chatbot-app
React	Production, enterprise	e2e-chatbot-app
Next.js	Modern production, SSR	e2e-chatbot-app-next

Why compatible? We follow the OpenAI API standard via MLflow AgentServer (/invocations endpoint).

Using official UIs:

# 1. Generate our backend
dbat generate chatbot my-bot

# 2. Clone official UI
git clone https://github.com/databricks/app-templates.git
cp -r app-templates/streamlit-chatbot-app my-bot/frontend

# 3. Point UI to backend (http://localhost:8000)
# 4. Deploy together!

Coming in v0.3.0: One-command integration!

dbat generate chatbot my-bot --ui=streamlit
# ✅ Backend + official Streamlit UI, pre-configured!

See UI Integration Guide for details.

---

API Documentation

Your agent comes with auto-generated OpenAPI documentation:

• Swagger UI: http://localhost:8000/docs
• OpenAPI JSON: http://localhost:8000/openapi.json
• Health Check: http://localhost:8000/health

Example API Call

import requests

response = requests.post(
    "http://localhost:8000/invocations",
    json={
        "input": [{"role": "user", "content": "Hello!"}],
        "stream": False
    }
)
# ResponsesAgent format: data.output[0].content[0].text
print(response.json()["output"][0]["content"][0]["text"])

Streaming Example

import requests
import json

response = requests.post(
    "http://localhost:8000/invocations",
    json={
        "input": [{"role": "user", "content": "Tell me a story"}],
        "stream": True
    },
    stream=True
)

for line in response.iter_lines():
    if line.startswith(b"data: "):
        data = line[6:].decode('utf-8')
        if data != "[DONE]":
            chunk = json.loads(data)
            # ResponsesAgent streaming format: chunk.content
            print(chunk.get("content", ""), end="", flush=True)

---

Features

Production-Ready

• ✅ FastAPI for performance and reliability
• ✅ OpenAPI schema for API documentation
• ✅ Health endpoints for monitoring
• ✅ Error handling and logging
• ✅ CORS configured for web UIs

Databricks-Native

• ✅ Auto-authentication with Databricks
• ✅ MLflow auto-tracing (track all LLM calls)
• ✅ Unity Catalog for data governance
• ✅ Model Serving integration
• ✅ Lakebase (PostgreSQL) for memory
• ✅ Vector Search for RAG

Developer Experience

• ✅ Configuration-driven (YAML, no code changes)
• ✅ CLI for quick testing
• ✅ Local development with hot-reload
• ✅ One-command deployment
• ✅ Comprehensive documentation

Standards-Based

• ✅ OpenAI API format (universal compatibility)
• ✅ Server-Sent Events (SSE) for streaming
• ✅ OpenAPI 3.0 schema
• ✅ REST best practices

---

Architecture

Your Application
┌─────────────────────────────────────────────────┐
│                                                 │
│  ┌─────────────┐        ┌──────────────────┐  │
│  │  Your UI    │◄──────►│  Agent Backend   │  │
│  │             │        │  (our toolkit)   │  │
│  │ - Streamlit │        │                  │  │
│  │ - React     │        │ - agent.py       │  │
│  │ - Gradio    │        │ - FastAPI server │  │
│  │ - Custom    │        │ - OpenAI API     │  │
│  └─────────────┘        └────────┬─────────┘  │
│                                  │             │
└──────────────────────────────────┼─────────────┘
                                   │
                    ┌──────────────┴──────────────┐
                    │   Databricks Platform       │
                    │                             │
                    │ - Model Serving (LLMs)      │
                    │ - MLflow (tracing)          │
                    │ - Lakebase (memory)         │
                    │ - Vector Search (RAG)       │
                    │ - Unity Catalog (data)      │
                    └─────────────────────────────┘

We provide the agent backend. You choose the UI.

---

Examples

Basic Chatbot

dbat generate chatbot hello-bot
cd hello-bot
python start_server.py

Assistant with Memory

dbat generate assistant support-bot
cd support-bot
# Configure Lakebase in databricks.yml
databricks bundle deploy

RAG-Powered Assistant

dbat generate assistant doc-bot --enable-rag
cd doc-bot

# Edit config.yaml:
# rag:
#   enabled: true
#   source: /Volumes/main/default/docs
#   backend: pgvector

databricks bundle deploy

Custom Model

dbat generate chatbot custom-bot
cd custom-bot

# Edit config.yaml:
# model:
#   endpoint: my-custom-endpoint
#   temperature: 0.9
#   max_tokens: 1000

python start_server.py

---

Requirements

• Python: 3.9+
• Databricks: Workspace access (for deployment)
• Model Serving: At least one LLM endpoint
• Optional:
  • Lakebase (for L2 memory)
  • Vector Search (for L2 RAG)
  • Unity Catalog Volumes (for RAG documents)

---

Documentation

• UI Integration Guide - Using official Databricks UI templates
• App Templates Compliance - Compatibility verification
• Upstream Sync Strategy - How we stay compatible

---

FAQ

Q: Do I need to use Databricks? A: For deployment, yes. For local development, you just need access to Databricks Model Serving endpoints.

Q: Can I use my own UI? A: Absolutely! Your agent backend follows the OpenAI API standard, so any OpenAI-compatible UI works.

Q: What about LangChain/LangGraph? A: Coming in v0.3.0+. For now, our agents use a simple, lightweight pattern. You can integrate LangChain yourself if needed.

Q: Is this production-ready? A: Yes! L1 (chatbot) is production-ready in v0.2.0. L2 (assistant) is in active testing.

Q: How do I switch models? A: Just edit config.yaml → model.endpoint. No code changes needed!

Q: Can I customize the agent logic? A: Yes! Edit agent.py - it's your code, do whatever you want.

Q: How do I add custom tools/functions? A: Modify the predict() method in agent.py to call your functions before/after the LLM.

---

Roadmap

v0.3.0 (Coming Soon)

• --ui=streamlit|gradio|react flag for one-command UI integration
• L3: API agents with custom tools
• L4: Multi-step workflows
• Template upgrade commands

v0.4.0+ (Future)

• L5: Multi-agent systems
• LangGraph integration
• Custom tool marketplace
• Performance benchmarking

---

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Areas we'd love help with:

• Additional UI framework integrations
• More example agents
• Documentation improvements
• Bug reports and feature requests

---

Philosophy

"On Top Of, Not Instead Of"

We don't reinvent wheels. We integrate official Databricks patterns:

• ✅ Official app-templates for UI
• ✅ OpenAI API standard
• ✅ MLflow for tracing
• ✅ FastAPI for servers
• ✅ Databricks services for infrastructure

We add:

• 🛠️ Scaffold generation (save time)
• ⚙️ Configuration management (no code changes)
• 📦 Pre-wired integrations (batteries included)
• 📚 Best practices (production-ready)

---

License

Apache 2.0

---

Support

• GitHub Issues: https://github.com/databricks/agent-toolkit/issues
• Documentation: See docs/ folder
• Examples: See examples/ folder

---

Built with ❤️ for the Databricks community

Start building agents today:

pip install databricks-agent-toolkit
dbat generate chatbot my-bot