Skip to main content

Python SDK

The Python is SDK built on top of the generated Python client library with the goal of providing excellent developer experience for building LLM security tools in Python. The SDK focusses on use-cases while hiding low level gRPC protocol details. This is the recommended way of using detoxio.ai APIs in Python based applications.

Installation​

python3 -m pip install detoxio \
  detoxio-api-protocolbuffers-python detoxio-api-grpc-python \
  --upgrade --extra-index-url https://buf.build/gen/python

Note: We need to install additional packages because as per PEP-440, we cannot have a package with direct dependencies outside the public index.

info

Install detoxio-api-protocolbuffers-pyi package for type hints and auto-completion in your IDE

LLM Security Scanner​

The LLM security scanner interface is used to scan an LLM for security vulnerabilities while maintaining necessary abstraction from LLM specific inference runtime. This ensures that detoxio SDK and APIs are decoupled from LLM specific frameworks such as TensorFlow, PyTorch etc. and can seamlessly integrate with any framework.

To get started, write a function that will take a detoxio.ai generated prompt, perform model specific inferencing and return the output generated by the model

from detoxio.scanner import LLMPrompt, LLMResponse

def model_adapter(prompt: LLMPrompt) -> LLMResponse:
  output = llm(input=prompt.content)  # llm(...) is a placeholder function
  return LLMResponse(content=output)

Use the LLMScanner class to start scanning an LLM using your model_adapter function

from detoxio.scanner import LLMScanner

scanner = LLMScanner(count=10)
result = scanner.start(prompt_handler=model_adapter)

Finally generate JSON report from the result produced by the scanner

from detoxio.reporting import JSONLinesLLMScanReport

report = JSONLinesLLMScanReport(file_path="/tmp/report.jsonl")
report.render(result)
note

The JSONL report generator will write JSON lines that are serialized PromptEvaluationResponse objects