Skip to main content

Introduction

Dhenara AI is an open source Python package for calling multiple LLM providers through one typed interface. It keeps the integration surface small: provider credentials, model selection, prompt and message formatting, streaming, tool use, reasoning, and structured output all flow through the same core client types.

Why Dhenara?​

  • Genuinely Open Source: Built from the ground up as a community resource, not an afterthought or internal tool
  • Unified API: Interact with different AI providers through a consistent interface
  • Type Safety: Built with Pydantic for robust type checking and validation
  • Easy Regeneration across Providers: With a unified Pydantic output and built-in prompt formatting, send output from a model to any other model easily
  • Streaming: First-class support for streaming responses along with accumulated responses similar to non-streaming responses
  • Async Support: Both synchronous and asynchronous interfaces for maximum flexibility
  • Centralized Resource Management: Keep provider credentials, APIs, and model endpoints in one place with ResourceConfig
  • Credential Security: Keep sensitive API keys and credentials separate from application code
  • Dynamic Model Selection: Switch between models and providers at runtime without reconfiguration
  • Provider Abstraction: Interact with foundation models regardless of which provider is serving them
  • Foundation Models: Pre-configured models with sensible defaults
  • Test Mode: Bring up your app with dummy responses for streaming and non-streaming generation
  • Cost/Usage Data: Derived cost and usage data along with responses, with optional charge for each model endpoint for commercial deployment
  • Community-Oriented Design: An architecture separating API credentials, models, and configurations for flexible deployment and scaling

Key Features​

  • Open Source and Extensible: Transparently designed codebase that encourages community contributions and extensions
  • Multiple Model Families: Work with OpenAI, Google, Anthropic, and other packaged model definitions through one response shape
  • Multiple API Providers: Support for direct OpenAI, Google Gemini Developer API, Anthropic, Google Vertex AI, Amazon Bedrock, and Microsoft OpenAI-compatible endpoints
  • Text and Image Generation: Generate text or images through the same interface
  • Streaming Support: Stream responses for better user experience
  • Accumulated Streaming Response: Process stream responses in the same way you do with non-streaming
  • File Integration: Easily incorporate files into your prompts
  • Cost Tracking: Monitor token usage and associated costs
  • Extensible Design: Add custom models, providers, or model configurations

Example Usage​

Here's a simple example using the public ResourceConfig credential flow:

from dhenara.ai import AIModelClient
from dhenara.ai.types import AIModelAPIProviderEnum, AIModelCallConfig, ResourceConfig

resource_config = ResourceConfig()
resource_config.load_from_file(credentials_file=None, init_endpoints=True)

endpoint = resource_config.get_model_endpoint(
    model_name="claude-haiku-4-5",
    api_provider=AIModelAPIProviderEnum.ANTHROPIC,
)
if endpoint is None:
    raise RuntimeError("No Anthropic endpoint configured for claude-haiku-4-5")

client = AIModelClient(
    model_endpoint=endpoint,
    config=AIModelCallConfig(max_output_tokens=300),
    is_async=False,
)

response = client.generate(
    prompt="Explain quantum computing in simple terms.",
    instructions=["Keep the answer under 120 words."],
)

assert response.chat_response
print(response.chat_response.text())

When credentials_file=None, ResourceConfig.load_from_file(...) resolves $DAI_SECRET_CONFIG_DIR/dai_credentials.yaml and falls back to /run/secrets/dai/dai_credentials.yaml if the environment variable is unset.

Next Steps​