External AI Functions

Build powerful AI/ML capabilities by connecting Databend with your own infrastructure. External functions let you deploy custom models, leverage GPU acceleration, and integrate with any ML framework while keeping your data secure.

Key Capabilities

Feature	Benefits
Custom Models	Use any open-source or proprietary AI/ML models
GPU Acceleration	Deploy on GPU-equipped machines for faster inference
Data Privacy	Keep your data within your infrastructure
Scalability	Independent scaling and resource optimization
Flexibility	Support for any programming language and ML framework

How It Works

1. Create AI Server: Build your AI/ML server using Python and databend-udf
2. Register Function: Connect your server to Databend with CREATE FUNCTION
3. Use in SQL: Call your custom AI functions directly in SQL queries

Databend Cloud Network Requirements

External functions use Arrow Flight over gRPC/HTTP2, not REST.

1. Expose the UDF server on HTTPS with a public TLS certificate and gRPC/HTTP2 support.
2. Open Support → Create New Ticket and add the hostname to the tenant UDF server allowlist.
3. If the firewall is locked down, allow Databend Cloud egress addresses on TCP 443.

CREATE FUNCTION fails with Unallowed UDF server address until the host is allowlisted, and it also verifies the remote schema, so the endpoint must be online.

Keep the local UDFServer gRPC listener private; terminate TLS at a gRPC load balancer:

Databend Cloud → HTTPS/HTTP2 :443 → UDFServer gRPC :8815

Example: Text Embedding Function

# Simple embedding UDF server demo
from databend_udf import udf, UDFServer
from sentence_transformers import SentenceTransformer

# Load pre-trained model
model = SentenceTransformer('all-mpnet-base-v2')  # 768-dimensional vectors

@udf(
    input_types=["STRING"],
    result_type="VECTOR(768)",
)
def ai_embed_768(text: str) -> list[float]:
    """Generate a 768-dimensional embedding for the input text."""
    embedding = model.encode(
        text or "",
        normalize_embeddings=True,
        show_progress_bar=False,
    )
    return embedding.tolist()

if __name__ == '__main__':
    print("Starting embedding UDF server on port 8815...")
    server = UDFServer("0.0.0.0:8815")
    server.add_function(ai_embed_768)
    server.serve()

-- Register the external function in Databend
CREATE OR REPLACE FUNCTION ai_embed_768 AS (STRING)
    RETURNS VECTOR(768)
    LANGUAGE python
    HANDLER = 'ai_embed_768'
    ADDRESS = 'https://your-ml-server.example.com';

-- Registration connects to the Flight endpoint and validates its schema.
SELECT VECTOR_DIMS(ai_embed_768('health check'));
-- 768

-- Use the custom embedding in queries
SELECT
    id,
    title,
    cosine_distance(
        ai_embed_768(content),
        ai_embed_768('machine learning techniques')
    ) AS similarity
FROM articles
ORDER BY similarity ASC
LIMIT 5;

Learn More

• External Functions Guide - Complete setup and deployment instructions
• Databend Cloud - Try external functions with a free account