smxadk 0.1.5

SyntaxMatrix Agent Development Kit for calling and deploying customer-owned SyntaxMatrix Agent Services.

smxADK

smxADK stands for SyntaxMatrix Agent Development Kit.

It is a lightweight Python SDK for calling a deployed SyntaxMatrix Agent Service from any Python project.

---

Installation

For local development:

pip install -e .

---

Basic Usage

from smxadk import SMXAgentClient

client = SMXAgentClient(
    base_url="https://your-agent-service-url"
)

response = client.chat(
    message="Explain RAG in two short sentences.",
    mode="expert",
)

print(response.answer)
print(response.usage.total_tokens)

---

---

Self-Hosting Model

smxADK is designed for self-hosted SyntaxMatrix deployments.

The client organisation owns and operates its own backend infrastructure:

Client Application
      ↓
smxADK
      ↓
Client-owned Agent Service
      ↓
Client-owned LiteLLM Proxy
      ↓
Client-owned Ollama / vLLM backends

This means:

• the client owns the backend URLs;
• the client pays for their own GPUs, CPUs, storage, and networking;
• the client controls their own data boundary;
• SyntaxMatrix provides the SDK, deployment tooling, templates, and framework.

smxADK should not be hardcoded to use SyntaxMatrix-owned infrastructure.
The caller must provide the deployed Agent Service URL:

from smxadk import SMXAgentClient

client = SMXAgentClient(
    base_url="https://client-owned-agent-service-url"
)

---

Deployment Configuration

Client organisations define their available model routes in:

smx_deployment.yaml

Example:

agent_service:
  supported_modes:
    - light
    - medium
    - heavy
    - expert
    - expert-heavy

models:
  light:
    provider: ollama
    model: your-light-model-name
    api_base: https://client-light-ollama-service-url

  medium:
    provider: ollama
    model: your-medium-model-name
    api_base: https://client-medium-ollama-service-url

  heavy:
    provider: ollama
    model: your-heavy-model-name
    api_base: https://client-heavy-ollama-service-url

  expert:
    provider: openai_compatible
    model: your-expert-model-name
    api_base: https://client-expert-vllm-service-url/v1    

  expert-heavy:
    provider: openai_compatible
    model: your-expert-heavy-model-name
    api_base: https://client-expert-heavy-vllm-service-url/v1

model: your-light-model-name model: your-medium-model-name model: your-expert-model-name

The client does not need to deploy every possible route.
They only declare the routes they actually have.

---

Generating Deployment Files

Run:

smxadk generate

This generates:

../llm-proxy/config.yaml
../agent-service/supported_modes.txt

The generated LiteLLM config maps each route to the client-owned backend URL.

The generated supported_modes.txt tells the Agent Service which modes it should accept.

---

Runtime Route Discovery

After deployment, applications can discover available routes:

from smxadk import SMXAgentClient

client = SMXAgentClient(
    base_url="https://client-owned-agent-service-url"
)

print(client.supported_modes())

Example output:

["light", "medium", "expert"]

If a caller requests a mode that is not supported, the Agent Service rejects it cleanly before calling LiteLLM.

---

Separation of Responsibilities

smx_deployment.yaml      → client backend routes and model catalogue
LiteLLM Proxy            → backend URL routing
Agent Service            → request validation and orchestration
smxADK                   → client SDK for application developers

The ADK only talks to the Agent Service.
It does not need to know the individual model backend URLs.

---

Health Check

from smxadk import SMXAgentClient

client = SMXAgentClient(
    base_url="https://your-agent-service-url"
)

health = client.health()

print(health.status)
print(health.supported_modes)

---

Streaming Chat

from smxadk import SMXAgentClient

client = SMXAgentClient(
    base_url="https://your-agent-service-url"
)

for chunk in client.stream_chat(
    message="Explain RAG in two short sentences.",
    mode="expert",
):
    print(chunk, end="", flush=True)

---

Supported Agent Service Endpoints

smxADK currently supports:

GET  /health
POST /chat
POST /chat/stream

---

Explicit Model Routing

The caller must explicitly choose the model route.

Example:

response = client.chat(
    message="Write a short summary.",
    mode="light",
)

Available modes depend on the deployed Agent Service.

Current example modes:

light
medium
heavy
expert
expert-heavy

---

Response Shape

client.chat() returns a ChatResponse object:

response.answer
response.mode
response.usage.prompt_tokens
response.usage.completion_tokens
response.usage.total_tokens

---

Project Structure

smx-adk/
├── pyproject.toml
├── README.md
└── smxadk/
    ├── __init__.py
    ├── client.py
    └── schemas.py

---

Design Goal

smxADK is designed to make deployed SyntaxMatrix Agent Services easy to plug into:

• FastAPI apps
• Dash apps
• Flask apps
• notebooks
• internal tools
• enterprise AI assistants
• SyntaxMatrix-based applications

---

Licence

Proprietary / SyntaxMatrix.