Why Fine-Tuning Quietly Breaks Your Embedding Index (and AlphaWiSE's Fix)
Fine-tune a CLIP-style model and the cross-modal alignment that made old queries work shifts underneath your index — recall erodes with no warning. AlphaWiSE interpolates two frozen checkpoints with one learned coefficient per tensor to keep both, at zero added inference cost.
Where this fits in Mixpeek
This is a Mixpeek operational hazard made concrete: change a collection's embedding model and every vector already in the index was written by the old model, so alignment drifts and recall silently degrades. The paper's fix (blend checkpoints) and its operational rule — pin embedding-model versions per collection — is exactly the discipline Mixpeek enforces so a model change never quietly breaks an existing index.
About this research
Fine-tune a CLIP-style model on new data and the cross-modal alignment that made old queries work shifts underneath your index; recall erodes with no dashboard warning. AlphaWiSE interpolates two frozen checkpoints with one learned coefficient per tensor, fitted on a small exemplar memory, and bakes out a single model with zero added inference cost. Also covers the operational fix: pin embedding-model versions per collection. Paper: arxiv.org/abs/2607.15094
Frequently asked questions
Go deeper
Put the research to work
Mixpeek turns video, images, audio, and documents in your object storage into searchable, timestamped results through one API — the retrieval stack these papers describe.
Search your own data, free