ONDC Merchant Catalog Guardrailing with CrewAI and NVIDIA NeMo
Small retail merchants face catalog ingestion errors and policy compliance violations when listing products on ONDC. Orchestrate a team of agents using CrewAI and NVIDIA NeMo Guardrails to sanitize and classify catalog items, achieving 99% safety policy compliance and under 2s per item processing time.
What you are building
Deploy a CrewAI collaborative agent workflow combined with NVIDIA NeMo Guardrails to enforce automated product listing safety and schema normalization.
How work is evaluated
Evaluates catalog item processing for safety policy compliance and output schema alignment.
Datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
safety_policy_enforcement
Verifies misleading claims are correctly flagged and suppressed.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
compliance_rate
Accuracy of safety policy classification across tests. • target: 0.99 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Learning goals
Build collaborative role-based agent workflows using CrewAI
Implement programmable guardrails using NVIDIA NeMo Guardrails
Normalize heterogeneous merchant catalog data to ONDC protocol schemas
Validate product claim policies against e-commerce consumer guidelines
Resources
200 raw D2C product catalog listings with varied formatting and policy violation edge cases.
Challenge setup
Evaluates catalog item processing for safety policy compliance and output schema alignment.
Challenge input
JSON containing raw_title, description, price, and category_claim.
CrewAI
Coordinates multi-agent role execution.
NVIDIA NeMo Guardrails
Enforces strict input/output policy constraints.
Evaluated output
JSON containing ondc_title, normalized_category, safety_passed (bool), and flag_reasons list.
- Verifies misleading claims are correctly flagged and suppressed.
- Accuracy of safety policy classification across tests. • target: 0.99 • range: 0-1
- Benchmark: E-Commerce Safety Guardrail Benchmark
- Compliance Rate target: 99%
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Shows the first task, up to three tools, and up to three evidence items.
View technical recipe
Configured tools
- CrewAI · Required
- RAI · Optional
- Nvidia · Optional
Evaluation contract
- safety_policy_enforcement · Weight 1
- compliance_rate · Weight 1
Metrics
- compliance_rate · Target 0.99
Benchmarks
- E-Commerce Safety Guardrail Benchmark
- security
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing raw_title, description, price, and category_claim.
Output: JSON containing ondc_title, normalized_category, safety_passed (bool), and flag_reasons list.
Execution harness: Python
Evaluation tools
- CrewAI · Agent Framework
- NVIDIA NeMo Guardrails · Safety Guardrails
Recipe freeze marker
No recipe freeze time is recorded.