Reproducible tests behind Stuga's videos and posts: the code, its captured output, and what it shows. Each folder runs on its own.
| Folder | Shows | Video |
|---|---|---|
| rerank-bench | 23 rerankers on the same candidates from MIRACL and BRIGHT StackOverflow: hit rate, latency and cost. | Claude Sonnet 5.5 vs Jev |
| bm25-score-leak | Raw BM25 scores over a shared index let a searcher estimate how many documents they cannot read contain a word. | Under the hood #1 |
| vector-search-empty | HNSW search with a permission filter can return nothing when readable matches exist, and how iterative scans and an exact scan fix it. | Under the hood #2 |
| embed-bench | EmbeddingGemma 2 and nine other embedding models on MIRACL, BRIGHT StackOverflow and C-MTEB, alone and fused with BM25 as Stuga retrieves, and why one similarity cutoff cannot suit them all. | EmbeddingGemma 2 benchmark; Under the hood #4 |
Code MIT. Data under each source's licence; names in the SQL tests are made up.