Distribution: 10K tests to make product research more reliable

Context

A major distribution player managed more than 10 product catalogs, each with tens of thousands of references and attributes (name, description, weight, units, packaging...). Its internal search engine (Elasticsearch) presented inaccurate and inconsistent results, slowing access to product information.

Challenge

• Uncertain index coverage: critical attributes sometimes not found
• Variable results: lack of consistency between exact searches and variants
• Common false positives related to fault tolerance
• Massive volume to test: thousands of requests to check

AI & Data Solution

• 400 audit requests to map indexing and identify missing fields
• 10K exact tests and 100K AI variations (modified words, spaces, punctuation) to assess robustness
• Semantic assessment by LLM to measure relevance and identify false positives
• Standardized prompt guide for reproducible tests

Results

✅ 100% of critical attributes verified
✅ Identified non-indexed fields and optimized mappings
✅ Major false positives fixed
✅ Test time reduced from several weeks to a few days

Conclusion

The company now has a reliable product search engine, with optimized precision and industrialized tests. A scalable methodology, applicable to any search engine requiring advanced quality control.