Skip to main content

​
The Synkro Pipeline

Synkro uses a multi-stage pipeline to generate high-quality training data: 
Document → Logic Map → Scenarios → Responses → Grade → Refine → Dataset

​
Stages

​
1. Document Ingestion

Your policy is parsed into raw text. Supported formats:
  • Plain text
  • PDF files
  • DOCX files
  • Markdown
  • URLs
from synkro.core.policy import Policy

# Multiple input methods
policy = Policy(text="Your policy text...")
policy = Policy.from_file("handbook.pdf")
policy = Policy.from_url("https://example.com/policy")

​
2. Logic Extraction

The policy is analyzed and structured rules are extracted into a Logic Map:
  • Each rule has an ID (R001, R002, etc.)
  • Conditions and actions identified
  • Dependencies between rules mapped
  • Categories assigned

​
3. Scenario Generation

Diverse test scenarios are created based on the Logic Map:
TypeDescription%
PositiveUser meets all criteria35%
NegativeUser violates one criterion30%
Edge CaseBoundary conditions25%
IrrelevantOutside policy scope10%

​
4. Response Generation

For each scenario, a policy-compliant assistant response is generated that:
  • Correctly applies the relevant rules
  • Cites specific policy sections
  • Handles edge cases appropriately

​
5. Grading

Each response is evaluated for:
  • Policy compliance
  • Accurate citations
  • Sound reasoning
  • Appropriate tone

​
6. Refinement

Failed responses are automatically refined (up to N iterations) until they pass quality checks.

​
7. Dataset Export

Passing traces are exported in your chosen format.

​
Pipeline Configuration

from synkro import create_pipeline, DatasetType
from synkro.models import Google

pipeline = create_pipeline(
    model=Google.GEMINI_25_FLASH,       # Fast generation
    grading_model=Google.GEMINI_25_PRO, # Quality grading
    dataset_type=DatasetType.CONVERSATION,
    max_iterations=3,                    # Refinement attempts
)

dataset = pipeline.generate(policy, traces=100)