Haystack Enterprise SDK

Experimental. This SDK is under active development. APIs and CLI commands may change without notice.

Python SDK and CLI for the Haystack Enterprise Platform.

Installation

# Install as a CLI tool
uv tool install haystack-enterprise-sdk

# Or add it as a dependency of your project
uv add haystack-enterprise-sdk

# pip works too
pip install haystack-enterprise-sdk

deploy, validate, and run load your pipeline in a subprocess using your project's own interpreter (an auto-detected venv, or --python), so the CLI environment does not need Haystack. Install the deploy extra only when the CLI environment doubles as the pipeline environment — that is, when there is no separate project venv to detect:

uv tool install "haystack-enterprise-sdk[deploy]"

To install unreleased changes from main:

uv tool install git+https://github.com/deepset-ai/haystack-enterprise-sdk.git

Usage

haystack-enterprise --help

Authentication

haystack-enterprise login
haystack-enterprise logout

Files

# Upload a folder to a workspace
haystack-enterprise upload ./my-files

# List files in a workspace
haystack-enterprise list-files

# Download files to your local machine
haystack-enterprise download

Pipelines

# Validate a local pipeline against the platform
haystack-enterprise validate ./pipeline.py

# Run a local pipeline in the platform sandbox (shows a spinner with the elapsed time while it waits)
haystack-enterprise run ./pipeline.py

# Transient failures (network errors, timeouts, 429, 5xx) are retried twice by default; 0 disables it.
# Config and input errors always fail immediately.
haystack-enterprise run ./pipeline.py --retries 0

# Deploy a local pipeline as a service deployment (reused if it exists, otherwise created serverless).
# Asks at most which socket receives the query and which is the main output -- and nothing at all when
# your socket names already say so. Prints the service's chat-completions endpoint once it is serving.
haystack-enterprise deploy ./pipeline.py my-service

# Create a managed (provisioned) service instead, with explicit sizing
haystack-enterprise deploy ./pipeline.py my-service --managed --cpu 2 --memory 4Gi

# Deploy and get a shareable prototype link (a chat UI). Because that UI routes through the pipeline's
# input/output mapping, --share is also what asks you to review the mapping.
haystack-enterprise deploy ./pipeline.py my-service --share

# Check the status of a service deployment
haystack-enterprise service-status my-service

Calling a deployed service

A deployed service is served over an OpenAI-compatible chat-completions endpoint, which deploy prints once the service is running:

curl -N https://api.cloud.deepset.ai/api/v1/workspaces/<workspace>/deployments/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "<workspace>/<deployment-id>", "messages": [{"role": "user", "content": "Hello"}]}'

The response is a server-sent-event stream of chat.completion.chunk objects. Any OpenAI client works — point its base_url at everything up to and including /deployments/v1.

Pass --verbose to any command for INFO/DEBUG logs, and <command> --help for its arguments.

Development

# Install uv if you don't have it
pip install uv

# Sync all dependencies (including dev dependencies)
uv sync --all-groups

# Run the CLI from source
uv run haystack-enterprise --help

See CONTRIBUTING.md for contribution guidelines.

Licenses

The SDK is licensed under Apache 2.0, see LICENSE.

Some bundled libraries are licensed under the MPL 2.0 license:

• tqdm for progress bars
• pathspec for pattern matching file paths
• certifi for validating trustworthiness of SSL certificates