| name | vec2text |
| description | vec2text: embedding-inversion library for reconstructing approximate text from sentence embeddings. Use when working with saved embedding tensors, privacy/inversion research, or AI/ML challenge workflows where you need to load a corrector model and invert strings or embeddings directly. |
| compatibility | Linux, macOS, Windows; Python 3; PyTorch and Hugging Face stack; GPU strongly recommended for larger runs |
| metadata | {"author":"AeonDave","version":"1.0"} |
vec2text
Use vec2text when embeddings are the artifact and text recovery is the question.
When to use vec2text
Use vec2text when you need to:
- invert sentence embeddings back into approximate text
- load a pretrained corrector for supported embedding families
- work from saved embedding tensors rather than original source text
- tune inversion quality with multiple refinement steps or beam search
Quick Start
import vec2text
corrector = vec2text.load_pretrained_corrector("gtr-base")
High-Value Workflows
Invert known strings through the matching embedder workflow
vec2text.invert_strings(
["example text"],
corrector=corrector,
num_steps=20,
sequence_beam_width=4,
)
Invert embeddings directly
import torch
embeddings = torch.load("embeddings.pt", map_location="cpu")
texts = vec2text.invert_embeddings(
embeddings=embeddings,
corrector=corrector,
num_steps=20,
)
print(texts)
Practical Notes
- Corrector choice must match the embedding family; mismatched models produce garbage.
num_steps improves refinement quality, while sequence_beam_width improves search at higher memory cost.
- The project supports both direct string inversion workflows and embedding-only inversion workflows.
Caveats
- Reconstruction is approximate, not guaranteed exact text recovery.
- Beam search raises memory usage quickly.
- Model support and pretrained aliases evolve; prefer current upstream docs over old examples.
Resources
No bundled scripts/, references/, or assets/.
Use the upstream vec2text README for current pretrained corrector aliases, API entry points, and embedding-family matching guidance.