Skip to main content

Python Client Library

The Python client library is auto-generated based on our API schema.

Installation​

Install the client library and its dependencies

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

Usage​

Setup authentication​

An API key is required to authenticate with the APIs. Refer to API authentication for more details. Once you have the API key, export it as an environment variable.

export DETOXIO_API_KEY=<your-api-key>

Setup gRPC channel​

import grpc
import os

token = grpc.access_token_call_credentials(os.getenv('DETOXIO_API_KEY'))
channel = grpc.secure_channel('api.detoxio.ai:443', grpc.composite_channel_credentials(grpc.ssl_channel_credentials(), token))

Setup API client​

import proto.dtx.services.prompts.v1.prompts_pb2 as dtx_prompts_pb2
import proto.dtx.services.prompts.v1.prompts_pb2_grpc as dtx_prompts_pb2_grpc
import proto.dtx.messages.common.llm_pb2 as dtx_llm_pb2

client = dtx_prompts_pb2_grpc.PromptServiceStub(channel)

Test for connectivity and credentials​

import google.protobuf.empty_pb2 as empty_pb2

client.Ping(empty_pb2.Empty())
warning

Authentication problems or rate limits will lead to an exception here

Generate a prompt​

res = client.GeneratePrompts(dtx_prompts_pb2.PromptGenerationRequest(count=1))
print(res.prompts[0].data.content)

Evaluate a prompt response​

The prompt evaluation API is meant to evaluate a response from an LLM under testing. You must use the generated prompt to perform an LLM inference to get it's output. The output from the LLM can subsequently used in this operation to identify vulnerabilities.

model_output = '<output from LLM under testing>'

# Create the request object for hydration
req = dtx_prompts_pb2.PromptEvaluationRequest()

# Pass the prompt generated earlier - res.prompts[0]
req.prompt.CopyFrom(res.prompts[0])
req.responses.extend([dtx_prompts_pb2.PromptResponse(message=dtx_llm_pb2.LlmChatIo(content=model_output))])

client.EvaluateModelInteraction(req)

Note on Datasets​

If you are generating datasets for model security testing and evaluation, you can use our threats taxonomy to categorize your dataset. This can help in generating useful reports across different threat categories.

import proto.dtx.messages.common.threat_pb2 as dtx_threat_pb2

Define an utility function to get name from constants

def dtx_get_threat_name(tid):
  return dtx_threat_pb2.ThreatCategory.DESCRIPTOR.values_by_number[tid].name

def dtx_get_threat_class(tid):
  return dtx_threat_pb2.ThreatClass.DESCRIPTOR.values_by_number[tid].name

Example categories

dtx_get_threat_class(dtx_threat_pb2.TOXICITY)
dtx_get_threat_name(dtx_threat_pb2.ABUSIVE_LANGUAGE)