Evaluation
6.3×
faster embedding with MiniLM than with BGE: 374.6 against 59.7 texts a second, on 5,000 texts
results fileTimed on 200K complaints on a T4 GPU in Google Colab: 20 fixed queries, five results each, one run, in milliseconds. Vector search includes encoding the query. There are no relevance labels, so this measures speed, not whether the results are right.
| method | p50 (ms) | p95 (ms) | p95 (ms) as a bar |
|---|---|---|---|
| Vector (FAISS) | 35 | 41 | |
| BM25 | 589 | 929 | |
| Hybrid (RRF) | 614 | 959 | |
| Reranked hybrid | 911 | 1,356 |
Stack
- FAISS
- Sentence-Transformers
- BERTopic
- Streamlit