Python SDK for Disseqt AI observability and validation. This package includes two SDKs:

1. Validation SDK (disseqt_sdk) - Validate LLM inputs, outputs, RAG, agentic behavior, and MCP security
2. Agentic SDK (disseqt_agentic_sdk) - Trace and monitor agentic AI workflows with OpenTelemetry-compatible spans

Documentation | API Reference | Examples

• Clean API: Single client.validate(request) method for all validators
• Type Safety: Full typing support with Python 3.10.14+
• Auto-Registration: Decorator-based validator registration system
• Normalized Responses: Consistent response format with dynamic others bag
• Domain-Specific Models: Module-scoped request types for each validation domain
• Enum-Driven: No raw strings in public API, everything uses enums

• OpenTelemetry Compatible: Standard span-based tracing
• Multiple Span Kinds: MODEL_EXEC, TOOL_EXEC, AGENT_EXEC, RAG_EXEC, MCP_EXEC, and custom kinds
• Automatic Batching: Efficient span batching and flushing
• Context Management: Thread-local context for nested traces
• Helper Functions: Decorators and utilities for easy integration

pip install disseqt-ai-sdk

pip install git+https://github.com/DisseqtAI/disseqt-python-sdk.git

For detailed installation instructions including virtual environments and troubleshooting, see INSTALL.md.

• ✅ You want to trace and monitor your agentic AI workflows
• ✅ You need observability for LLM calls, tool calls, and agent actions
• ✅ You want to track span hierarchies (parent-child relationships)
• ✅ You need OpenTelemetry-compatible tracing
• ✅ You want to visualize workflows in the Disseqt dashboard
• ✅ You're building RAG systems and need RAG_EXEC spans
• ✅ You're using MCP (Model Context Protocol) and need MCP_EXEC spans

• ✅ You want to validate LLM inputs/outputs for safety and quality
• ✅ You need to check for toxicity, bias, prompt injection, etc.
• ✅ You want to evaluate RAG grounding (context relevance, faithfulness)
• ✅ You need to assess agentic behavior (goal accuracy, tool call accuracy)
• ✅ You want MCP security validation (data leakage, insecure output)
• ✅ You need composite scoring across multiple metrics

Trace your agentic AI workflows for observability and monitoring:

from disseqt_agentic_sdk import DisseqtAgenticClient, start_trace
from disseqt_agentic_sdk.enums import SpanKind
from disseqt_agentic_sdk.api.helpers import trace_llm_call, trace_tool_call

# Initialize client
client = DisseqtAgenticClient(
    api_key="your-api-key",
    project_id="proj_456",
    service_name="my-service",
    endpoint="http://localhost:8080/v1/traces"
)

# Create a trace with multiple spans
with start_trace(client, name="agent_workflow") as trace:
    # Agent execution span
    with trace.start_span("agent_execution", SpanKind.AGENT_EXEC) as agent_span:
        agent_span.set_agent_info("my_agent", "agent_001")
        
        # LLM call span
        with trace.start_span("llm_call", SpanKind.MODEL_EXEC) as llm_span:
            llm_span.set_model_info("gpt-4", "openai")
            llm_span.set_messages(
                input_messages=[{"role": "user", "content": "Hello"}],
                output_messages=[{"role": "assistant", "content": "Hi there!"}]
            )
            llm_span.set_token_usage(input_tokens=10, output_tokens=5)
        
        # Tool call span
        with trace.start_span("api_call", SpanKind.TOOL_EXEC) as tool_span:
            tool_span.set_tool_info("get_weather", "call_001")
            tool_span.set_attribute("tool.input.city", "Paris")

# Trace automatically sent when exiting the 'with' block
client.shutdown()

For RAG (Retrieval Augmented Generation) workflows, use RAG_EXEC span kind:

with start_trace(client, name="rag_workflow") as trace:
    with trace.start_span("rag_retrieval", SpanKind.RAG_EXEC) as rag_span:
        # Set context and messages for RAG validation
        rag_span.set_attribute("agentic.input.context", "Retrieved context...")
        rag_span.set_messages(
            input_messages=[{"role": "user", "content": "What is AI?"}],
            output_messages=[{"role": "assistant", "content": "AI is..."}]
        )

For MCP (Model Context Protocol) workflows, use MCP_EXEC span kind:

with start_trace(client, name="mcp_workflow") as trace:
    with trace.start_span("mcp_execution", SpanKind.MCP_EXEC) as mcp_span:
        mcp_span.set_messages(
            input_messages=[{"role": "user", "content": "Query"}],
            output_messages=[{"role": "assistant", "content": "Response"}]
        )
        mcp_span.set_attribute("mcp.protocol.version", "1.0")

You can use custom span kinds by passing any string:

# Custom span kind
with trace.start_span("data_processing", "DATA_PROCESSING") as span:
    span.set_attribute("processing.type", "batch")

# Or with decorator
from disseqt_agentic_sdk.api.helpers import trace_function

@trace_function(client, kind="CUSTOM_OPERATION")
def my_function():
    return "result"

For more Agentic SDK examples, see the agentic examples directory.

The Composite Score Evaluator combines multiple validators for comprehensive LLM output evaluation:

from disseqt_sdk import Client
from disseqt_sdk.models.composite_score import CompositeScoreRequest
from disseqt_sdk.validators.composite.evaluate import CompositeScoreEvaluator

# Initialize client
client = Client(project_id="your_project_id", api_key="your_api_key")

# Simple composite evaluation
evaluator = CompositeScoreEvaluator(
    data=CompositeScoreRequest(
        llm_input_query="What is the capital of France?",
        llm_output="The capital of France is Paris.",
    )
)

result = client.validate(evaluator)
overall = result.get("overall_confidence", {})
print(f"Score: {overall.get('score')}, Label: {overall.get('label')}")

For advanced usage with custom weights and thresholds (see full example):

evaluator = CompositeScoreEvaluator(
    data=CompositeScoreRequest(
        llm_input_query="What are the differences between men and women in parenting?",
        llm_input_context="Research shows that both men and women can be effective parents.",
        llm_output="Women are naturally better at nurturing children than men.",
        evaluation_mode="binary_threshold",
        weights_override={
            "top_level": {
                "factual_semantic_alignment": 0.50,
                "language": 0.25,
                "safety_security_integrity": 0.25,
            },
            "submetrics": {
                "factual_semantic_alignment": {
                    "factual_consistency": 0.70,
                    "answer_relevance": 0.05,
                    "conceptual_similarity": 0.05,
                    "compression_score": 0.05,
                    "rouge_score": 0.05,
                    "cosine_similarity": 0.02,
                    "bleu_score": 0.02,
                    "fuzzy_score": 0.02,
                    "meteor_score": 0.04,
                },
                "language": {
                    "clarity": 0.40,
                    "readability": 0.30,
                    "response_tone": 0.30,
                },
                "safety_security_integrity": {
                    "toxicity": 0.30,
                    "gender_bias": 0.15,
                    "racial_bias": 0.15,
                    "hate_speech": 0.20,
                    "data_leakage": 0.15,
                    "insecure_output": 0.05,
                },
            },
        },
        labels_thresholds_override={
            "factual_semantic_alignment": {
                "custom_labels": ["Low Accuracy", "Moderate Accuracy", "High Accuracy", "Excellent Accuracy"],
                "label_thresholds": [0.4, 0.65, 0.8],
            },
            "language": {
                "custom_labels": ["Poor Quality", "Fair Quality", "Good Quality", "Excellent Quality"],
                "label_thresholds": [0.25, 0.5, 0.7],
            },
            "safety_security_integrity": {
                "custom_labels": ["High Risk", "Medium Risk", "Low Risk", "Minimal Risk"],
                "label_thresholds": [0.6, 0.8, 0.95],
            },
        },
        overall_confidence={
            "custom_labels": ["Low Confidence", "Moderate Confidence", "High Confidence", "Very High Confidence"],
            "label_thresholds": [0.4, 0.55, 0.8],
        },
    )
)
result = client.validate(evaluator)

from disseqt_sdk import Client, SDKConfigInput
from disseqt_sdk.models.input_validation import InputValidationRequest
from disseqt_sdk.models.output_validation import OutputValidationRequest
from disseqt_sdk.models.agentic_behaviour import AgenticBehaviourRequest
from disseqt_sdk.validators.input.safety import ToxicityValidator
from disseqt_sdk.validators.output.accuracy import FactualConsistencyValidator
from disseqt_sdk.validators.agentic_behavior.reliability import TopicAdherenceValidator

# Initialize client
client = Client(project_id="proj_123", api_key="key_xyz")

# Input validation
toxicity_validator = ToxicityValidator(
    data=InputValidationRequest(prompt="What do you think about politics?"),
    config=SDKConfigInput(threshold=0.5),
)
result = client.validate(toxicity_validator)
print(result)

# Output validation
fact_validator = FactualConsistencyValidator(
    data=OutputValidationRequest(response="The Eiffel Tower is in Paris and was built in 1889."),
    config=SDKConfigInput(threshold=0.6),
)
result = client.validate(fact_validator)
print(result)

# Agentic behaviour validation
topic_validator = TopicAdherenceValidator(
    data=AgenticBehaviourRequest(
        conversation_history=["user: Tell me about deep learning.", "agent: I like pizza."],
        tool_calls=[],
        agent_responses=["I like pizza."],
        reference_data={"expected_topics": ["machine learning", "neural networks", "artificial intelligence", "deep learning"]},
    ),
    config=SDKConfigInput(threshold=0.8),
)
result = client.validate(topic_validator)
print(result)

For Agentic SDK (tracing/observability) examples, see the agentic examples directory:

• example.py - Complete example with multiple spans using context managers
• example_without_with.py - Manual span management without with statements
• ai_consultant_agent.py - Full AI agent integration example

For Validation SDK examples, see the examples directory:

• example.py - Comprehensive examples of all validator types (Input, Output, Agentic, MCP, RAG)
• example_composite_score.py - Composite Score Evaluator with multi-metric evaluation
• verify_installation.py - Installation verification script

Each example includes:

• Complete working code
• API configuration
• Error handling
• Output interpretation

For full API documentation, visit docs.disseqt.ai.

All validators return a normalized response:

{
  "data": {
    "metric_name": "topic_adherence_evaluation",
    "actual_value": 0.4571191966533661,
    "actual_value_type": "float",
    "metric_labels": ["Always Off-Topic"],
    "threshold": ["Fail"],
    "threshold_score": 0.8,
    "others": { "...": "dynamic" }
  },
  "status": { "code": "200", "message": "Success" }
}

Safety & content moderation for user inputs:

• ToxicityValidator - Detects toxic content in input text
• BiasValidator - Detects general bias in input
• InputPromptInjectionValidator - Detects prompt injection attempts
• IntersectionalityValidator - Analyzes intersectional bias
• RacialBiasValidator - Detects racial bias
• GenderBiasValidator - Detects gender bias
• PoliticalBiasValidator - Detects political bias
• SelfHarmValidator - Detects self-harm content
• ViolenceValidator - Detects violent content
• TerrorismValidator - Detects terrorism-related content
• SexualContentValidator - Detects sexual content
• HateSpeechValidator - Detects hate speech
• NSFWValidator - Detects NSFW content
• InvisibleTextValidator - Detects hidden/invisible text attacks
• ChildSafetyValidator - Detects child-safety risks in input

Per-project intent guardrails (configurable block/allow lists):

• IntentGuardValidator - Blocks disallowed intents (block list; response enforcement "blocking")
• IntentComplianceValidator - Flags intents outside the allow list (enforcement "advisory")

Quality Metrics:

• FactualConsistencyValidator - Checks factual accuracy of output
• AnswerRelevanceValidator - Measures answer relevance to the question
• ClarityValidator - Evaluates clarity of response
• CoherenceValidator - Measures logical coherence
• ConceptualSimilarityValidator - Measures conceptual similarity
• CreativityValidator - Evaluates creativity of response
• DiversityValidator - Measures response diversity
• GrammarCorrectnessValidator - Checks grammar correctness
• NarrativeContinuityValidator - Evaluates narrative flow
• ReadabilityValidator - Measures readability level
• ResponseToneValidator - Analyzes response tone

Safety & Bias Detection:

• OutputToxicityValidator - Detects toxic content in output
• OutputBiasValidator - Detects bias in output
• OutputGenderBiasValidator - Detects gender bias in output
• OutputRacialBiasValidator - Detects racial bias in output
• OutputPoliticalBiasValidator - Detects political bias in output
• OutputHateSpeechValidator - Detects hate speech in output
• OutputNSFWValidator - Detects NSFW content in output
• OutputSelfHarmValidator - Detects self-harm content in output
• OutputSexualContentValidator - Detects sexual content in output
• OutputTerrorismValidator - Detects terrorism content in output
• OutputViolenceValidator - Detects violent content in output
• OutputIntersectionalityValidator - Detects intersectional bias in output
• OutputChildSafetyValidator - Detects child-safety risks in output

Security:

• OutputDataLeakageValidator - Detects data leakage in output
• OutputInsecureOutputValidator - Detects insecure output patterns

Intent Guardrails:

• OutputIntentGuardValidator - Blocks disallowed intents in the model's output (block list; enforcement "blocking")
• OutputIntentComplianceValidator - Flags output intents outside the allow list (enforcement "advisory")

Scoring Metrics:

• BleuScoreValidator - Calculates BLEU score
• RougeScoreValidator - Calculates ROUGE score
• MeteorScoreValidator - Calculates METEOR score
• CosineSimilarityValidator - Calculates cosine similarity
• FuzzyScoreValidator - Calculates fuzzy matching score
• CompressionScoreValidator - Measures compression ratio

Validators for Retrieval-Augmented Generation systems:

• ContextRelevanceValidator - Validates context relevance
• ContextRecallValidator - Measures context recall
• ContextPrecisionValidator - Measures context precision
• ContextEntitiesRecallValidator - Measures entity recall from context
• NoiseSensitivityValidator - Evaluates noise sensitivity
• ResponseRelevancyValidator - Measures response relevancy to context
• FaithfulnessValidator - Measures faithfulness to source context

Validators for AI agent evaluation:

• TopicAdherenceValidator - Ensures agents stay on topic
• ToolCallAccuracyValidator - Measures tool call accuracy
• ToolFailureRateValidator - Tracks tool failure rates
• PlanOptimalityValidator - Evaluates plan optimality
• AgentGoalAccuracyValidator - Measures goal achievement accuracy
• IntentResolutionValidator - Evaluates intent resolution
• PlanCoherenceValidator - Measures plan coherence
• FallbackRateValidator - Tracks fallback rates

Security validators for Model Context Protocol:

• McpPromptInjectionValidator - Detects prompt injection attempts
• DataLeakageValidator - Detects data leakage
• InsecureOutputValidator - Detects insecure output patterns

Multi-metric evaluation:

• CompositeScoreEvaluator - Combines multiple validators for comprehensive scoring

• ClassifyValidator - Classifies content into themes/categories

All validators require a configuration object:

config = SDKConfigInput(
    threshold=0.8,
    custom_labels=["Low Risk", "Medium Risk", "High Risk"],
    label_thresholds=[0.3, 0.7]
)

client = Client(
    project_id="your_project_id",
    api_key="your_api_key",
    base_url="https://api.disseqt.ai/realtime-validations",  # Default
    timeout=30  # Default timeout in seconds
)

Each validation domain has its own request model:

• InputValidationRequest: For input validation (prompt, optional context/response)
• OutputValidationRequest: For output validation (response)
• RagGroundingRequest: For RAG validation (prompt, context, response)
• AgenticBehaviourRequest: For agentic validation (conversation_history, tool_calls, etc.)
• McpSecurityRequest: For MCP security (prompt, optional context/response)
• CompositeScoreRequest: For composite scoring (llm_input_query, llm_output, evaluation_mode, weights)
• ThemesClassifierRequest: For theme classification (text, return_subthemes, max_themes)

The SDK raises HTTPError for API failures:

from disseqt_sdk.client import HTTPError

try:
    result = client.validate(validator)
except HTTPError as e:
    print(f"API Error {e.status_code}: {e.message}")
    print(f"Response: {e.response_body}")

# Clone and setup
git clone https://github.com/DisseqtAI/disseqt-python-sdk.git
cd disseqt-python-sdk
uv sync

# Install pre-commit hooks
uv run pre-commit install

# Run tests with coverage
uv run pytest -q --cov=disseqt_sdk --cov-report=term-missing

# Run linting
uv run ruff check .
uv run black --check .
uv run mypy

1. Create validator file in appropriate domain directory
2. Subclass the correct base validator class
3. Add @register_validator decorator
4. Import in domain's __init__.py
5. Add tests

Example:

from dataclasses import dataclass
from ...enums import ValidatorDomain, InputValidation
from ...registry import register_validator
from ..base import InputValidator

@register_validator(
    domain=ValidatorDomain.INPUT_VALIDATION,
    slug=InputValidation.NEW_VALIDATOR.value,
    path_template="/api/v1/sdk/validators/{domain}/{validator}",
)
@dataclass(slots=True)
class NewValidator(InputValidator):
    def __post_init__(self) -> None:
        object.__setattr__(self, "_domain", ValidatorDomain.INPUT_VALIDATION)
        object.__setattr__(self, "_slug", InputValidation.NEW_VALIDATOR.value)

Proprietary - Copyright (c) 2024 Disseqt AI Limited. All rights reserved.

For support and licensing inquiries, contact: support@disseqt.ai