Load vLLM data to DuckDB

vLLM is a high-throughput inference engine that provides an OpenAI-compatible REST API for serving large language models. Everything needed to build a working vLLM → DuckDB pipeline is on this page: the API's base URL, authentication, endpoints, pagination and incremental field — plus a prompt that hands the whole job to your coding agent.

---

Build your vLLM to DuckDB pipeline

Paste this prompt into Claude, Codex, or Cursor. The agent does the rest.

PromptCopy prompt
Run uvx dlthub-init@latest to build a pipeline from vLLM to DuckDB and run it on dltHub

That scaffolds a dltHub workspace and installs the dltHub AI harness — the project rules, the secrets-management skill, and the dlt MCP server your agent needs to work safely. From there it reads the vLLM API, proposes the endpoints to load, then writes, runs and validates the pipeline while you review rather than type. Credentials are inspected through MCP tools, so your agent never reads secrets.toml itself. How the LLM-native workflow works →

Prefer to write it yourself? Every fact the agent uses is below.

---

vLLM API at a glance

These values come from the vLLM API reference — the authoritative source if anything here looks out of date.

---

How do I authenticate with the vLLM API?

When the --api-key flag (or VLLM_API_KEY environment variable) is configured, the server enforces Bearer token authentication in the 'Authorization' header for protected endpoints. The header must follow the format 'Authorization: Bearer <YOUR_API_KEY>'.

1. Get your credentials

vLLM is a self-hosted inference server and does not provide a web-based dashboard for managing API credentials. Credentials are set by the administrator during server startup. To configure an API key, provide it via the --api-key command-line argument (e.g., vllm serve MODEL_NAME --api-key YOUR_SECRET_KEY) or set the VLLM_API_KEY environment variable before launching the server. Clients must then include this key in the Authorization: Bearer YOUR_SECRET_KEY header for all requests to protected /v1 endpoints.

2. Add them to .dlt/secrets.toml

Copy
[sources.vllm_source]
api_key = "your_api_key_here"

dlt reads this file automatically at runtime. With the harness, the setup-secrets skill prompts you for the values and never handles the raw credential in chat. For production, see setting up credentials with dlt.

---

What vLLM data can I load into DuckDB?

These are the vLLM endpoints dlt can load into DuckDB:

---

How do I load only new vLLM records?

The vLLM API reference does not document a timestamp or sequence field for these endpoints, so there is nothing to advertise here as verified. Pick a field from the endpoints table above that increases with every write, then set it as the cursor_path.

Copy
{"name": "models", "endpoint": {
    "path": "v1/models",
    # Replace with a field that increases on every write.
    "incremental": {"cursor_path": "REPLACE_ME", "initial_value": "2024-01-01T00:00:00Z"},
}}

On the first run dlt loads everything from initial_value; on every run after that it requests only what changed and appends with write_disposition="merge" if you set a primary key. See incremental loading.

---

What does the generated vLLM pipeline look like?

A standard dlt REST API pipeline — the same code you would write by hand, loading /v1/chat/completions and /v1/completions from the vLLM API into DuckDB:

Copy
import dlt
from dlt.sources.rest_api import RESTAPIConfig, rest_api_resources

@dlt.source
def vllm_source(api_key=dlt.secrets.value):
    config: RESTAPIConfig = {
        "client": {
            "base_url": "http://localhost:8000/v1",
            "auth": {"type": "bearer", "token": api_key},
        },
        "resources": [
            {"name": "models", "endpoint": {"path": "v1/models", "data_selector": "data"}},
            {"name": "health", "endpoint": {"path": "health"}}
        ],
    }
    yield from rest_api_resources(config)


def load_vllm_to_duckdb() -> None:
    pipeline = dlt.pipeline(
        pipeline_name="vllm_pipeline",
        destination="duckdb",
        dataset_name="vllm_data",
    )
    load_info = pipeline.run(vllm_source())
    print(load_info)


if __name__ == "__main__":
    load_vllm_to_duckdb()

Run it with python vllm_pipeline.py. The agent iterates on this until it loads cleanly — you review and approve, rather than write it from scratch.

---

How do I query vLLM data in DuckDB?

dlt creates one table per resource. Query the loaded data with Python or SQL — or ask your agent to, through the MCP server's execute_sql_query tool.

Python (pandas DataFrame):

Copy
import dlt

data = dlt.pipeline("vllm_pipeline").dataset()
df = data.models.df()
print(df.head())

SQL:

Copy
SELECT * FROM vllm_data.models LIMIT 10;

See querying your data with dataset and exploring it in marimo notebooks.

---

How do I deploy the vLLM to DuckDB pipeline in production?

The pipeline runs locally, which is ideal for prototyping and one-off analysis. When you need it on a schedule, monitored on every load, and shared with your team, deploy the same dlt code on the dltHub platform — no infrastructure to maintain. The prompt above already ends with "run it on dltHub", so your agent can take it there directly.

• Deploy & schedule — run the pipeline as a managed job with automatic retries.
• Monitor — observable job queues, alerting, and load metrics for every run.
• Transform — promote raw vLLM loads into governed, documented models.
• Visualize & share — explore data in notebooks and publish live dashboards instead of static screenshots.

Book a demo →

---

What other destinations can I load vLLM data to?

dlt loads into any of these — only the destination argument changes:

Set dlt.pipeline(destination="snowflake") and add credentials in .dlt/secrets.toml. On the dltHub platform the same pipeline runs against a managed Iceberg lakehouse. See the full destinations list.

---

Next steps

• Ready for production? Book a demo of the dltHub platform →
• dlt LLM-native workflow docs
• REST API source reference
• Join the dlt community on Slack