Getting Started with ADRI

PyPI • GitHub

Stop AI agents from breaking on bad data in under 5 minutes

Step 1 – Install & Bootstrap

pip install adri

# Create ADRI project folders and sample data
adri setup --guide

The guided setup creates ADRI/dev and ADRI/prod folders plus a sample invoice dataset so you can see ADRI working before wiring it into your own data.

Step 2 – Generate a Standard from Good Data

adri generate-standard examples/data/invoice_data.csv \
  --output examples/standards/invoice_data_ADRI_standard.yaml --guide

This command profiles your clean dataset and writes a YAML standard that encodes required fields, value ranges, and freshness expectations. The --guide flag explains what ADRI is doing at each step.

Step 3 – Validate New Data with the Standard

adri assess examples/data/test_invoice_data.csv \
  --standard examples/standards/invoice_data_ADRI_standard.yaml --guide

What you should see:

• Allowed ✅ when data complies with the generated standard
• Blocked ❌ with a summary of failed checks when the test data violates the standard

🛡️ ADRI Protection: BLOCKED ❌
📊 Quality Score: 62.4/100 (Required: 80.0/100)
⚠️ Issues: amount must be > 0, status must be lowercase paid/pending/...

When the score is above the threshold ADRI prints ALLOWED ✅ so you know the data is safe for your agent.

Step 4 – Protect a Function

from adri import adri_protected

@adri_protected(standard="invoice_data_standard", data_param="invoice_rows")
def process_invoices(invoice_rows):
    # Your existing agent logic
    return ai_agent(invoice_rows)

Key parameters:

• standard – Name of the YAML standard (without .yaml) you generated or stored in ADRI/**/standards
• data_param – The function argument that contains the data you want ADRI to validate
• min_score (optional) – Override the required score (default comes from config)
• on_failure (optional) – Choose how to respond: "raise", "warn", or "continue"

See Core Concepts for protection modes and the five dimensions.

Change Failure Handling

# Warn-only: log issues but keep running (great for pilots)
@adri_protected(standard="invoice_data_standard", data_param="invoice_rows", on_failure="warn")

# Selective mode: continue but mark the run
@adri_protected(standard="invoice_data_standard", data_param="invoice_rows", on_failure="continue")

# Strict mode: raise ProtectionError when data fails validation (default)
@adri_protected(standard="invoice_data_standard", data_param="invoice_rows")

Common Framework Patterns

# LangChain
@adri_protected(standard="invoice_data_standard", data_param="invoice_rows")
def langchain_support_agent(invoice_rows):
    chain = prompt | model | parser
    return chain.invoke(invoice_rows)

# CrewAI
@adri_protected(standard="invoice_data_standard", data_param="invoice_rows")
def crewai_market_analysis(invoice_rows):
    crew = Crew(agents=[analyst], tasks=[analysis_task])
    return crew.kickoff(inputs=invoice_rows)

# Generic Python function
@adri_protected(standard="invoice_data_standard", data_param="invoice_rows")
def your_agent_function(invoice_rows):
    return your_ai_framework(invoice_rows)

Configure ADRI (Optional)

Create ADRI/config.yaml (or edit the generated one) to customise defaults:

adri:
  project_name: "invoice-agents"
  default_environment: "development"

  protection:
    default_min_score: 80
    default_failure_mode: "raise"   # raise | warn | continue
    cache_duration_hours: 1
    auto_generate_standards: true
    verbose_protection: false

  environments:
    development:
      paths:
        standards: "./ADRI/dev/standards"
        assessments: "./ADRI/dev/assessments"
        training_data: "./ADRI/dev/training-data"
      protection:
        default_min_score: 75
        default_failure_mode: "warn"
    production:
      paths:
        standards: "./ADRI/prod/standards"
        assessments: "./ADRI/prod/assessments"
        training_data: "./ADRI/prod/training-data"
      protection:
        default_min_score: 90
        default_failure_mode: "raise"

What ADRI Validates

ADRI automatically checks five dimensions:

• Validity – Correct formats (emails, dates, enumerations)
• Completeness – Required fields populated
• Consistency – Data types and formats aligned
• Plausibility – Realistic value ranges
• Freshness – Data recency and relevance

Logs & Audit Trail

Every assessment captures:

• Console summary (success/failure + score)
• JSON reports in ADRI/**/assessments
• Optional CSV audit trail when enabled
• Enterprise mode streaming to Verodat MCP (see Adoption Journey)

Troubleshooting

ValueError: Could not find data parameter – Make sure data_param matches the name of the argument in your protected function.

Standard file not found – Confirm the path you pass to --standard exists. Use adri list-standards to see registered files.

Validation too strict? – Lower the threshold temporarily:

@adri_protected(standard="your_data_standard", data_param="invoice_rows", min_score=60)

See what ADRI is checking – Inspect the YAML:

adri show-standard invoice_data_standard

Next Steps

1. Work through a framework playbook – Copy/paste solutions for LangChain, CrewAI, LlamaIndex, and more.
2. Review the FAQ – Understand logging, standards, and roadmap.
3. Explore real examples – Business scenarios and walkthroughs.
4. Follow the Adoption Journey – Know when to turn on Verodat-managed supply.

---

Need help? Open an issue on GitHub or join the discussion.

Next steps

Framework PlaybooksCore ConceptsAPI ReferenceAdoption Journey