Open-source embedflow migrates embedding models with zero downtime, skipping costly re-embedding

Potential_Low_1183 · reddit · 2026-09-08

A researcher released embedflow, tackling the huge cost of upgrading embedding models in RAG pipelines: re-embedding 1 billion docs with Qwen Embed 8B on an H100 takes about 108 days.

The method is simple: instead of rebuilding the index, take K documents from the old index and rerank them with the new model. With sufficient K, retrieval quality matches the target model's native index. The author ran 63 migrations (up to 1M docs); the best result was upgrading Qwen 4B to 8B with just 50 documents matching native retrieval, entirely avoiding expensive backfill.

It works with Qdrant and is installable via pip install embedflow; code is on GitHub.

Original post →

Open-source embedflow migrates embedding models with zero downtime, skipping costly re-embedding

More from coding & agent