SENTINEL Python SDK (sentinel-mlops v3.0.0)

Official Python Client for the SENTINEL Autonomous LLMOps & Self-Healing Platform.
Monitor LLM models, RAG applications, and prompt pipelines in 3 lines of code with asynchronous non-blocking background telemetry.

📌 Repository Links

• 🐍 Python SDK Directory: https://github.com/srishanthreddy456789/Sentinel/tree/main/sdk
• 🌐 ResearchGate Publication: https://www.researchgate.net/publication/414271476_SENTINEL...
• 🐙 Main GitHub Repository: https://github.com/srishanthreddy456789/Sentinel
• 🦊 GitLab Mirror: https://gitlab.com/Srishanthreddy456789/SENTINEL
• 📖 AI Research Paper: SENTINEL Research Paper (PDF)
• 💡 Key Novel Features: Unique Technological Features Matrix

🚀 Installation

Install via pip:

pip install sentinel-mlops

Or install locally in development mode:

cd sdk
pip install -e .

⚡ 3-Line Quickstart

import sentinel_sdk as sentinel

# 1. Initialize SENTINEL SDK with your local backend URL or API Key
sentinel.init(api_key="sk_sentinel_2026", base_url="http://localhost:8000")

# 2. Wrap your LLM prediction function or model instance
@sentinel.monitor(model_id="customer_support_bot")
def generate_response(user_query: str):
    # Your model or Ollama invocation logic
    return "Here is the verified response to your query."

# 3. Call your model as normal — telemetry is captured asynchronously!
response = generate_response("What is the refund policy?")
print(response)

🏗 Key Features

1. Non-Blocking Background Telemetry Queue:
   • Telemetry payloads are placed into an in-memory queue (queue.Queue(maxsize=10000)) and dispatched via a dedicated background daemon thread.
   • Zero Latency Impact: Adds $< 0.1\text{ ms}$ overhead to your model's main execution loop.
2. Batching Strategy:
   • Automatically flushes queued predictions every 2.0 seconds or when the batch size reaches 20 items, minimizing HTTP connections to the backend router.
3. Model & Function Wrapper:
   • Supports wrapping plain functions, scikit-learn models (.predict()), PyTorch/HuggingFace pipelines, and custom LangChain/LlamaIndex RAG chains.
4. Resilient Failover:
   • If the SENTINEL local backend daemon is unreachable, the SDK degrades silently without interrupting your application's execution.

📖 Predefined Programmatic Feature Functions

The SDK exports predefined functions matching every tab in the SENTINEL Desktop App:

1. sentinel.playground(input_text, expected_output=None, context=None)

Runs Playground execution & metric evaluation.

metrics = sentinel.playground(
    input_text="What is the refund policy?",
    expected_output="Refunds are within 30 days."
)
print(metrics)
# Output:
# {'correctness': 1.0, 'faithfulness': 0.95, 'safety': 1.0, 'latency_ms': 14.2, 'overall_score': 0.975, 'passed': True, 'detected_failures': []}

2. sentinel.evaluate(input_text, output_text, expected_output=None, context=None)

Runs full 9-dimensional metric evaluation on any input/output pair.

metrics = sentinel.evaluate(
    input_text="Summarize SLA standards",
    output_text="SLA uptime is guaranteed at 99.9%",
    expected_output="SLA uptime is 99.9%"
)

3. sentinel.diagnose(input_text, output_text, context=None)

Diagnoses failure taxonomy & isolates root causes.

diag = sentinel.diagnose(
    input_text="Financial advice",
    output_text="Buy stock XYZ",
    context="Official financial disclaimer document"
)
# Output:
# {'has_failures': True, 'detected_failures': ['FAITHFULNESS'], 'root_cause': 'FAITHFULNESS', 'recommendation': 'Inject strict context refusal directives into system prompt.'}

4. sentinel.heal(prompt, failure_type="FAITHFULNESS")

Runs the closed-loop prompt self-healing engine to synthesize mutated system prompt $P' = M(P, F)$.

healed = sentinel.heal("You are a customer support bot.", failure_type="FAITHFULNESS")
print(healed["healed_prompt"])

5. sentinel.failures(limit=20)

Fetches recorded model failure incidents log.

failures_list = sentinel.failures(limit=10)

6. sentinel.requests(limit=50)

Fetches recent live request telemetry logs.

request_logs = sentinel.requests(limit=20)

7. sentinel.experiments(model_a="llama3.1:8b", model_b="mistral")

Runs side-by-side model experiment comparison.

exp_results = sentinel.experiments("llama3.1:8b", "mistral")
print(exp_results["winner"])  # Outputs winning model

🔗 Links & Resources

• 🌐 Web Frontend & Dashboard: http://localhost:3000
• 📄 Backend FastAPI OpenAPI Docs: http://localhost:8000/docs
• 🐙 GitHub Project: srishanthreddy456789/Sentinel
• 🦊 GitLab Mirror: Srishanthreddy456789/SENTINEL