SDK v0.4.0

Updated AImodel flow

SDK v0.3.3

Login fixes and more

SDK v0.3.2

Login fixes and more

SDK v0.3.1

Bug fixes and improvements

SDK v0.3.0

Ability to login with username and password

SDK v0.2.0

DataSpace SDK v0.2.0 Release Notes

Release Date: November 12, 2025
Type: Minor Release (Feature Addition)

---

🚀 New Features

AI Model Execution Support

Added comprehensive support for executing AI models directly through the SDK:

• call_model(model_id, input_text, parameters) - Synchronous model execution
• call_model_async(model_id, input_text, parameters) - Asynchronous execution (placeholder for Celery integration)

Supported Providers:

• OpenAI (GPT-3.5, GPT-4, etc.)
• Llama (Ollama, Together AI, Replicate, Custom endpoints)
• HuggingFace (local and remote inference)
• Custom API endpoints

Example:

from dataspace_sdk import DataSpaceClient

client = DataSpaceClient(base_url="https://api.dataspace.example.com")
client.login(keycloak_token="your_token")

result = client.aimodels.call_model(
    model_id="uuid",
    input_text="Hello, world!",
    parameters={"temperature": 0.7, "max_tokens": 100}
)

print(result["output"])  # Model response

---

HuggingFace Integration

Added comprehensive HuggingFace model support with new fields:

Field	Type	Description
hf_use_pipeline	Boolean	Use Pipeline inference API
hf_auth_token	String	Authentication token for gated models
hf_model_class	Enum	Model head specification (8 variants)
hf_attn_implementation	String	Attention function (e.g., flash_attention_2)
framework	Enum	PyTorch or TensorFlow

Supported Model Classes:

• AutoModelForCausalLM
• AutoModelForSeq2SeqLM
• AutoModelForSequenceClassification
• AutoModelForNextSentencePrediction
• AutoModelForMultipleChoice
• AutoModelForTokenClassification
• AutoModelForQuestionAnswering
• AutoModelForMaskedLM

---

CRUD Operations

Full model lifecycle management:

# Create
model = client.aimodels.create({
    "name": "my-gpt-model",
    "display_name": "My GPT Model",
    "provider": "OPENAI",
    "provider_model_id": "gpt-4",
    "model_type": "TEXT_GENERATION",
    "is_public": True
})

# Update
client.aimodels.update(model_id, {
    "status": "ACTIVE",
    "is_public": False
})

# Delete
client.aimodels.delete_model(model_id)

---

📊 API Enhancements

New Endpoints

Method	Endpoint	Description
POST	/api/aimodels/{id}/call/	Execute model inference
POST	/api/aimodels/{id}/call-async/	Async model execution
POST	/api/aimodels/	Create new model
PATCH	/api/aimodels/{id}/	Update model
DELETE	/api/aimodels/{id}/	Delete model

Response Format

Standardized response for model execution:

{
  "success": true,
  "output": "Model generated response...",
  "latency_ms": 245.67,
  "status_code": 200,
  "provider": "OpenAI",
  "raw_response": {
    "choices": [...],
    "usage": {...}
  }
}

Error Response:

{
  "success": false,
  "error": "Error message...",
  "latency_ms": 123.45,
  "status_code": 500
}

---

🔐 Security & Access Control

Organization-Based Access

Models can now be restricted to organization members:

# Only organization members can access private models
if not model.is_public and model.organization:
    # Check user's organization membership
    if user not in model.organization.members:
        return 403 Forbidden

Authentication

• Full Keycloak integration
• Automatic token refresh
• Secure API key storage with encryption

---

🏗️ Backend Services

ModelAPIClient

Features:

• Multiple authentication methods (Bearer, API Key, Custom)
• Request template system with placeholder replacement
• Provider-specific request/response handling
• Automatic retry with exponential backoff (3 attempts)
• Endpoint statistics tracking

Supported Patterns:

# OpenAI format
{"model": "gpt-4", "messages": [...]}

# Llama/Ollama format
{"model": "llama2", "prompt": "...", "stream": false}

# Custom template
{"input": "{input}", "params": {...}}

ModelHFClient

Features:

• Pipeline-based inference
• Manual model loading with custom configurations
• Automatic GPU/CPU device selection
• Batch processing support
• Comprehensive error handling

---

📦 New Enums

AIModelProvider

OPENAI = "OPENAI"
LLAMA_OLLAMA = "LLAMA_OLLAMA"
LLAMA_TOGETHER = "LLAMA_TOGETHER"
LLAMA_REPLICATE = "LLAMA_REPLICATE"
LLAMA_CUSTOM = "LLAMA_CUSTOM"
CUSTOM = "CUSTOM"
HUGGINGFACE = "HUGGINGFACE"  # NEW

AIModelFramework

PYTORCH = "pt"
TENSORFLOW = "tf"

HFModelClass

CAUSAL_LM = "AutoModelForCausalLM"
SEQ2SEQ_LM = "AutoModelForSeq2SeqLM"
SEQUENCE_CLASSIFICATION = "AutoModelForSequenceClassification"
# ... and 5 more

---

🔧 Technical Improvements

Type Safety

• Full mypy compatibility
• Comprehensive type annotations
• Proper handling of Union types
• Type ignores for third-party libraries

Error Handling

• Graceful degradation for missing endpoints
• Detailed error messages
• Proper exception hierarchy
• Async/sync context handling

Performance

• Connection pooling with httpx
• Retry logic with exponential backoff
• Endpoint statistics tracking
• Async support throughout

---

📋 Migration Guide

From v0.1.0 to v0.2.0

✅ No Breaking Changes - Fully backward compatible

New Capabilities:

1. Model Execution

   # Old: Not available
   # New: Execute models directly
   result = client.aimodels.call_model(model_id, "input text")

2. CRUD Operations

   # Old: Read-only access
   # New: Full CRUD support
   client.aimodels.create({...})
   client.aimodels.update(id, {...})
   client.aimodels.delete_model(id)

3. HuggingFace Support

   # Old: Limited provider support
   # New: Full HuggingFace integration
   model = client.aimodels.get_by_id_graphql(hf_model_id)
   print(model["hfModelClass"])

---

🔗 Dependencies

New Dependencies

httpx>=0.28.1          # Async HTTP client
tenacity>=9.1.2        # Retry logic
torch>=2.9.0           # PyTorch (optional, for HuggingFace)
transformers>=4.57.1   # HuggingFace models (optional)
nest-asyncio>=1.6.0    # Nested event loop support

Installation

# Basic installation
pip install dataspace-sdk>=0.2.0

# With HuggingFace support
pip install dataspace-sdk[huggingface]>=0.2.0

---

🧪 Testing

Unit Tests

from unittest.mock import patch

@patch('dataspace_sdk.resources.aimodels.AIModelClient.call_model')
def test_model_execution(mock_call):
    mock_call.return_value = {
        "success": True,
        "output": "Test response"
    }
    
    result = client.aimodels.call_model("uuid", "test")
    assert result["success"] is True

Integration Tests

# Requires running DataSpace backend
client = DataSpaceClient(base_url="http://localhost:8000")
client.login(keycloak_token="test_token")

models = client.aimodels.list_all()
assert len(models) > 0

---

🐛 Bug Fixes

• Fixed type annotations for mypy compatibility
• Improved error handling for missing endpoints
• Corrected async/sync execution context handling
• Fixed response extraction for various providers
• Resolved Union type issues with AnonymousUser

---

🎯 Use Cases

1. Centralized Model Registry - Manage all AI models in one place
2. Multi-Provider Execution - Unified interface for different providers
3. Access Control - Organization-based model access
4. Monitoring - Track model usage and performance
5. A/B Testing - Compare different models easily

---

🔮 Roadmap (v0.3.0)

• Celery-based async execution
• Streaming response support
• Batch inference API
• Model versioning
• Cost tracking and analytics
• Rate limiting per organization
• Model performance metrics

---

📖 Documentation

• API Reference
• Authentication Guide
• Model Execution Guide
• Migration Guide

---

🙏 Acknowledgments

Special thanks to all contributors who helped make this release possible!

---

📦 Installation

pip install dataspace-sdk==0.2.0

Requirements:

• Python >= 3.11
• Django >= 5.2
• PostgreSQL (for backend)

---

🔗 Links

• PyPI: https://pypi.org/project/dataspace-sdk/
• GitHub: https://github.com/CivicDataLab/DataSpaceBackend
• Documentation: https://docs.dataspace.example.com
• Issues: https://github.com/CivicDataLab/DataSpaceBackend/issues

---

Full Changelog: v0.1.0...v0.2.0

Initial Release

Supports CRUD Operations on Dataset, UseCase, Collaboratives
Ability to create sharts for datafiles.