Dobby SDK

Lightweight multi-provider LLM SDK with streaming and tool support.

Installation

pip install dobby-sdk

# Or from GitHub
pip install git+https://github.com/TYNYBAY/dobby-sdk.git

# With uv
uv add dobby-sdk

Quick Start

from dobby import AgentExecutor, OpenAIProvider
from dobby.types import UserMessagePart, TextPart, TextDeltaEvent

provider = OpenAIProvider(model="gpt-4o", api_key="sk-...")
executor = AgentExecutor(provider="openai", llm=provider)

messages = [UserMessagePart(parts=[TextPart(text="Hello!")])]

async for event in executor.run_stream(messages):
    match event:
        case TextDeltaEvent(delta=delta):
            print(delta, end="")

Vertex AI Credentials Setup

VertexAIProvider authenticates via a google.auth.credentials.Credentials object — no API key needed. Store the service account key as a single env var (works the same locally and in prod, no key file on disk):

1. Download the service account key JSON (GCP Console → IAM & Admin → Service Accounts → Keys), minify it to one line, and add it to your .env (or your prod secret manager's env injection) as GOOGLE_APPLICATION_CREDENTIALS_JSON:

   GOOGLE_APPLICATION_CREDENTIALS_JSON='{"type":"service_account","project_id":"...","private_key":"-----BEGIN PRIVATE KEY-----\n...\n-----END PRIVATE KEY-----\n", ...}'
   

   Single quotes, one line — keep \n as the literal two-character escape inside the JSON string, not a real newline. Never commit this value.

2. Grant the service account the roles/aiplatform.user role on the project (IAM console), and enable the Vertex AI API:

   gcloud services enable aiplatform.googleapis.com --project=YOUR_PROJECT_ID
   

3. That's it — VertexAIProvider picks up GOOGLE_APPLICATION_CREDENTIALS_JSON itself when credentials isn't passed explicitly, no manual credential-building needed:

   import asyncio
   
   from dobby.providers import VertexAIProvider
   from dobby.types import UserMessagePart, TextPart
   
   async def main():
       provider = VertexAIProvider(
           model="meta/llama-3.1-405b-instruct-maas",
           project="your-project-id",
           location="us-central1",
           scopes=["https://www.googleapis.com/auth/cloud-platform"],
       )
       result = await provider.chat([UserMessagePart(parts=[TextPart(text="Hello!")])])
       print(result.parts)
   
   asyncio.run(main())
   

   uv run --env-file .env python examples/vertexai_example.py
   

   Note: scopes is required here — a service-account credential (whether from GOOGLE_APPLICATION_CREDENTIALS_JSON or a key file) has no implicit scope; omitting it fails with invalid_scope: Invalid OAuth scope or ID token audience provided.

On GCP itself (Cloud Run, GKE, Compute Engine): skip all of the above — attach a service account to the runtime and let ADC resolve it automatically via the metadata server. No key material to manage at all. The env-var approach above is only needed off-GCP (other clouds, local dev without gcloud auth application-default login).

Other auth styles — a service-account key file, service-account impersonation, or plain ADC — are shown as runnable variants in build_provider() in examples/vertexai_example.py, with the full reference in docs/providers/vertexai.md.

Features

• Multi-provider: OpenAI, Azure OpenAI, Anthropic (direct, Azure AI Foundry), Gemini (Developer API), Vertex AI Model Garden
• Streaming: Real-time token streaming with typed events
• Tools: Dataclass-based tools with auto-generated schemas
• Context injection: Pass runtime context to tools via Injected[T]
• Structured output: Pydantic model validation for agent responses

Documentation

See docs/ for detailed documentation:

• Getting Started
• Message Types
• Providers
• Tools
• AgentExecutor
• Vector Stores
• Retrievers

License

MIT