Enrichment should add a traceable, reviewable interpretation to existing authorized data. It must not fabricate facts or infer sensitive traits. Separate source facts, deterministic derived values, model suggestions, and human-approved fields.
Define the Enrichment Contract
For a public business-description dataset:
input: record_id, approved public description, source URL/version
output: category from allowed taxonomy, short summary, evidence span
never infer: owner identity, revenue, religion, ethnicity, political view, creditworthiness
state: SUGGESTED until accepted
retention: source and decision evidence only as required
Deduplicate input records and cache only when tenant, source version, prompt, model, and schema match. A source update invalidates enrichment.
Define reviewer disagreement categories such as wrong taxonomy, insufficient evidence, harmful inference, stale source, or unclear input. Report them by prompt/model/source version with minimum sample sizes. This turns corrections into an evaluation program instead of letting a workflow silently accumulate low-quality labels.
Build the Pipeline
trigger → authorization/minimization → deterministic validation
→ AI structured suggestion → schema/policy validation
→ confidence/evidence gate → human review or accepted suggestion
→ write with provenance → metrics
Do not let the model browse arbitrary URLs. Retrieve approved content through a controlled fetcher with destination and size limits.
Worked Example
A Karachi marketplace categorizes seller-supplied public product descriptions into 12 catalog groups. Fifty labelled fixtures include ambiguous and bilingual copy. The model returns category, evidence quote limited to the source, and needs_human.
Unknown products route to review; the model cannot invent material, brand, warranty, or origin. Approved results store source hash, model/prompt/schema version, reviewer, and timestamp. A seller correction replaces the suggestion and triggers reevaluation.
Failure Cases to Diagnose
- Enrichment overwrites source fact: store separate SUGGESTED field.
- No provenance: link input/version/model/reviewer.
- Sensitive inference marketed as personalization: prohibit and audit.
- Cached result crosses client/updated source: include tenant/version keys.
- Model evidence not in input: reject.
- Human corrections never feed evaluation: record labelled errors safely.
🇵🇰 Pakistan Angle
Test Roman Urdu and local product vocabulary with authorized examples. Do not assume a name, neighborhood, language, or price reveals customer class or protected traits.
For ecommerce, enrichment can suggest category or missing-field questions but cannot certify authenticity, halal status, safety, legal compliance, or warranty without authoritative evidence and qualified review.
Hands-On Exercise
- Write permitted/prohibited fields.
- prepare 30 labelled synthetic/public fixtures.
- build structured suggestion pipeline.
- add evidence, cache versioning, and human review.
- measure errors by category/language.
Completion Rubric
- Sources are authorized and minimized.
- Suggestions never overwrite authority.
- Sensitive inference is prohibited.
- Evidence/provenance are validated.
- Cache is tenant/version safe.
- Corrections and evaluator metrics exist.
Sources
Key takeaway: enrichment is a versioned suggestion with source evidence and review; it must never silently become a fabricated or sensitive business fact.