AI GitHub
vec2text for studying embedding inversion
Explore methods that reconstruct text from embeddings and test privacy assumptions.
Key Takeaways
- The research code uses learned models to predict and refine text given an embedding.
- Use the project's supported inversion and embedding models with synthetic data.
- Work with synthetic sentences and a supported embedding model.
- Results are specific to the embedding model and attack setup.
How it works
The research code uses learned models to predict and refine text given an embedding. Its experiments show why transformed representations can still reveal source information.
What you need
Use the project's supported inversion and embedding models with synthetic data. Matching the embedding pipeline is part of reproducing the reconstruction experiment.
What to try first
Work with synthetic sentences and a supported embedding model. Compare reconstruction quality for common phrases and unique strings.
Try a synthetic record with a unique token or identifier. Inspect whether it is recovered, approximated or replaced with a more common string.
Where to look for failures
Results are specific to the embedding model and attack setup. Keep private data out of the experiment and avoid treating low reconstruction quality as proof of privacy.
Related Franklin pages
Separate encoder capability, retrieval quality and privacy assumptions. Each needs a different test, even when all three experiments start from the same set of embeddings.
- Study an encoder architecture for tasks such as classification, retrieval and token labeling. ModernBERT for encoder-based NLP
- Evaluate embeddings on documented retrieval, classification and other benchmark tasks. MTEB for comparing embedding models across tasks
For the role of the encoder in a search stack, read Franklin’s model guide. Embedding models
Project documentation
Try it for your workflow
Explore the project’s documentation for setup and usage, or save this page to a collection for later.
Read the project documentation ↗