The Retrieval Bench
Controlled evaluations of EmbeddingGemma 2 against BM25, other embedding baselines, and simple hybrid-search configurations with reproducible benchmarks.
Why This Blog Exists: Testing Retrieval Systems Instead of Repeating Model Claims
This inaugural post introduces Semantic Workshop, a platform dedicated to independent, reproducible field notes on Embed...
EmbeddingGemma 2 vs. BM25: The First Reproducible Comparison
This inaugural comparison evaluates EmbeddingGemma 2 against BM25 on a small, noisy support dataset of 1,200 documents a...
How to Build a Small Retrieval Evaluation Set Without Pretending It Is Universal
This post details how to build a small, practical retrieval evaluation set using 1,200 documents and 80 queries instead ...
Recall@10, MRR, and NDCG: Three Metrics, One Concrete Example
This post breaks down three core retrieval evaluation metrics—Recall@10, MRR, and NDCG@10—using a single concrete query ...
The Benchmark Is the Conversation Starter, Not the Verdict
This post establishes the core operating philosophy for Semantic Workshop, arguing that retrieval benchmarks should serv...