Datasets:
metadata
license: mit
task_categories:
- feature-extraction
language:
- en
tags:
- embeddings
- vector-database
- migration
- retrieval
- ms-marco
- beir
pretty_name: Embedding Migration Results
size_categories:
- n<1K
Embedding Migration Results
Experimental results from testing whether vector databases can be migrated to new embedding models without re-embedding the entire corpus.
Overview
We embedded ~1M MS MARCO passages with 6 embedding models across 3 dimensionalities (768, 1024, 2560), trained linear translators between every pair of embedding spaces, and measured recall@10 ratio (translated / native ceiling).
Models
| Model | Dim | Prefix |
|---|---|---|
| intfloat/e5-base-v2 | 768 | query/passage |
| BAAI/bge-base-en-v1.5 | 768 | query only |
| Alibaba-NLP/gte-base-en-v1.5 | 768 | none |
| nomic-ai/nomic-embed-text-v1.5 | 768 | search_query/search_document |
| Qwen/Qwen3-Embedding-0.6B | 1024 | instruction format |
| Qwen/Qwen3-Embedding-4B | 2560 | instruction format |
Files
| File | Evaluations | Description |
|---|---|---|
msmarco_procrustes.json |
30 | Orthogonal Procrustes / least-squares on MS MARCO 1M |
msmarco_ridge.json |
30 | Ridge regression on MS MARCO 1M |
beir_scifact_procrustes.json |
30 | Procrustes/LS evaluated on BEIR SciFact |
beir_scifact_ridge.json |
30 | Ridge evaluated on BEIR SciFact |
beir_fiqa_procrustes.json |
30 | Procrustes/LS evaluated on BEIR FiQA |
beir_fiqa_ridge.json |
30 | Ridge evaluated on BEIR FiQA |
beir_nfcorpus_procrustes.json |
30 | Procrustes/LS evaluated on BEIR NFCorpus |
beir_nfcorpus_ridge.json |
30 | Ridge evaluated on BEIR NFCorpus |
fewshot_curve.json |
84 | Few-shot learning curve (50-5000 training examples) |
relative_repr.json |
30 | Relative representations baseline (Moschella et al. 2023) |
Total: 354 evaluations
Key Results
- 50 paired examples is enough for 95%+ native performance on compatible model pairs
- Prefix mismatch (not dimension mismatch) is the dominant failure mode
- Ridge regression rescues prefix-mismatched pairs (E5→GTE: 0.090 → 0.814)
- Cross-domain generalization holds for same-family pairs, degrades 10-30% for mismatched pairs
- Relative representations fail at retrieval scale (negative result)
Schema
Each JSON file contains an array of evaluation records. Common fields:
{
"source": "model/name",
"target": "model/name",
"method": "procrustes|ridge|relative_repr",
"src_dim": 768,
"tgt_dim": 768,
"recall@10_translated": 0.862,
"recall@10_native_target": 0.885,
"ratio": 0.974,
"train_size": 5000
}
Citation
If you use these results, please link the GitHub repository: https://github.com/allison-stack/embedding-migration
License
MIT