offCanada/Compliments_Product_Grouping_Pipeline

Open Food Facts Product Identity Resolution Pipeline

Overview

This repository contains an experimental product identity resolution pipeline built on Open Food Facts product data.

The source dataset contains duplicate, incomplete, inconsistently named, and potentially conflicting product records. The objective is not simply to find products with similar names, but to determine which records represent the same underlying product so that duplicate records can eventually be grouped and incomplete information can be safely enriched from other records in the same validated product group.

For example, if two records represent the same product but one record contains more complete nutrition or product information, the validated product group may later be used to fill missing attributes in the less complete record.

The pipeline is designed to remain brand-agnostic so that the same approach can be extended beyond the initial data to other grocery brands.

Problem Definition

Product similarity and product identity are treated as different problems.

For example:

• SOYA SAUCE and Sauce Soya may represent the same underlying product despite differences in word order.
• Different package sizes may still represent the same underlying product.
• Salted Butter and Unsalted Butter are highly similar names but represent different product variants.
• Products from the same category may have high embedding similarity without representing the same product.

The pipeline therefore combines deterministic normalization, lexical and semantic candidate generation, probabilistic record linkage, and semantic identity reasoning rather than relying on a single similarity score.

Pipeline Architecture

The current implementation is divided into three notebooks.

Notebook 1 — Data Preparation (EDA_OpenDB)

Purpose: Prepare a consistent product-level dataset for downstream matching.

Main activities include:

• selecting and validating relevant product fields
• cleaning and normalizing product names
• resolving and standardizing brand information
• standardizing quantity information
• preparing nutrition fields
• creating normalized matching keys and
• validating the prepared dataset.

Output: opendb_cleaned_products.parquet

This file is the standardized input used by subsequent notebooks.

Notebook 2 — Candidate Generation & Evidence (Candidate_Pairs)

Purpose: Generate high-recall candidate product pairs and collect multiple independent similarity signals.

Candidate generation intentionally casts a wider net than the final identity decision. A candidate pair means that two records are sufficiently related to investigate. It does not mean they are the same product.

Evidence includes:

• exact normalized-name matching
• word-reordering evidence
• RapidFuzz/string similarity
• word-level TF-IDF similarity
• character-level TF-IDF similarity
• sentence-embedding similarity
• shared and differing product-name tokens
• quantity/package evidence and
• nutrition similarity/conflict evidence.

Sentence embeddings are used for semantic similarity and candidate discovery, not as proof of product identity. Semantically related products can receive high embedding similarity while still representing different products.

Output: opendb_candidate_pairs_evidence.parquet

This file contains candidate record pairs together with the evidence required for downstream identity resolution.

Notebook 3 — Product Identity Resolution (Splink_Comparisons)

Purpose: Determine whether candidate records represent the same underlying product.

Notebook 3 uses a hybrid approach consisting of:

1. Splink probabilistic record linkage
2. Structural identity checks
3. LLM-assisted semantic product identity resolution for ambiguous cases

Splink evaluates product-name and nutrition evidence probabilistically. However, experiments showed that probabilistic similarity alone cannot fully resolve product identity.

For example, highly similar variants such as:

Churned Unsalted Butter Sticks vs. Churned Salted Butter Sticks

can receive very high probabilistic similarity because most words and nutritional characteristics agree, despite representing different formulations.

Conversely, structurally equivalent names with missing supporting information, such as:

SOYA SAUCE vs. Sauce Soya

may receive a low overall probabilistic score.

For this reason, Splink probability is treated as one source of evidence rather than the final match decision.

Structural Identity Resolution

High-confidence deterministic relationships can bypass semantic review.

Examples include:

• exact normalized product names and
• identical meaningful tokens appearing in a different order.

These relationships are considered structurally strong identity evidence.

Pairs containing additional, removed, or changed meaningful tokens are treated more cautiously because those differences may represent flavours, formulations, ingredients, textures, varieties, or simply incomplete source data.

LLM-Assisted Product Identity Resolution

Ambiguous candidate relationships are prepared for semantic review.

Only candidates with sufficient lexical, semantic, embedding, or probabilistic evidence are included. These thresholds determine whether a relationship is worth reviewing. They are not product-match thresholds.

Ambiguous relationships are represented as a graph:

• product records are nodes
• candidate relationships are edges and
• connected components form semantic review batches.

A semantic review batch is not a product group.

Its purpose is only to provide enough context for an LLM to reason about related products together.

For example, a batch could contain:

• Cherry Jelly Powder
• Cherry Jello Powder
• Lime Jelly Powder
• Orange Jelly Powder

The semantic reviewer may determine that the first two records represent one underlying product while the lime and orange variants should remain separate.

The LLM is asked to reason about product identity rather than general semantic similarity and to remain conservative when evidence is insufficient.

Semantic Review Evidence

The semantic reviewer can receive:

• product names
• package quantity
• nutrition information
• Splink probability
• shared tokens
• left/right-only tokens
• lexical similarity
• TF-IDF similarities
• embedding similarity and
• nutrition agreement/conflict information.

No individual signal independently establishes product identity.

LLM Independence

The semantic-review layer is designed to be provider-independent.

The notebook prepares standardized prompts and structured JSON responses while the actual LLM call is isolated behind a single interface. This allows different hosted or local models to be evaluated without modifying the rest of the identity-resolution pipeline.

All LLM responses must be validated before downstream use to ensure that:

• every expected record is returned
• records are not duplicated
• no record IDs are invented and
• unresolved records can remain explicitly uncertain.

LLM Experimentation

Two LLM models were evaluated during development:

1. Gemini 3.6 Flash:

Gemini Flash was initially used to validate the semantic-review architecture. While it was successful in identifying and processing a few semantic batches, it also exposed practical issues including:

• API Quota limits
• Temporary Server errors
• Occasionally structurally inconsistent JSON outputs.

2. GPT-OSS-120B through Groq:

GPT-OSS produced comparable results on overlapping test batches and was used to complete the semantic review of all review components. The implementation includes API retry handling, quota-aware stopping, checkpoint/resume functionality, JSON validation, record-ID validation, and special compact-prompt handling for unusually large review batches.

One review component that exceeded the normal request size was successfully processed after reducing the semantic prompt by approximately 14.2% without splitting the connected component.

Output: opendb_product_identity_links.parquet

Notebook 4 - Product Grouping

Metric	Result
Source records	1,322
Resolved product groups	1,110
Multi-record groups	154
Singleton groups	956
Records participating in multi-record groups	366
Product identity representations consolidated	212
Product identity consolidation rate	~16.0%
Grouping participation rate	~27.7%
Singleton rate	~72.3%
Average multi-record group size	~2.38

Current Pipeline

Raw Open Food Facts Data
          |
          v
   Data Preparation
          |
          v
 Candidate Generation + Evidence Creation
          |
          v
 Probabilistic Matching
        Splink
          |
          v
 Structural Identity Checks
          |
          +----------------------+
          |                      |
          v                      v
 Deterministic Match       Ambiguous Cases
                                  |
                                  v
                         LLM Identity Review
                                  |
                                  v
                       Validated Identity Links