Offer normalization

#Objective

Determine whether offers from different sources are comparable enough to appear in the same result set.

#Initial strategy

Use deterministic extraction whenever possible.

Examples:

• normalize units;
• parse "500 ml", "0.5 L", "500ml";
• standardize categories;
• separate toppings or modifiers;
• normalize currency values.

Use model-assisted extraction only when deterministic parsing is insufficient.

#Principle

An LLM may help extract structure, but it should not silently decide equivalence.

The system should preserve:

• original source title;
• extracted attributes;
• normalization method;
• confidence;
• reasons for match or rejection.

#Open research problem

Equivalence is category-specific. Matching logic for açaí may not generalize to pizza, meals, or branded products.