Our take
- Activity
- 19
- Community
- 14
- Issues
- 21
- Pull requests
- 20
+0 stars in the last 30 days
- Fewer contributors than 86% of the projects we track
Project history
Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.
Stars
43Forks
3Change since Nov 17, 2026: 0 stars·0 forks
Overview
NUDGE is a Python tool for adapting pre-trained document embeddings to a retrieval task. Rather than retraining an embedding model, it moves the data embeddings under constraints, using training queries and their known relevant records to improve similarity-based retrieval.
It is suited to retrieval and RAG practitioners who already have embeddings and labeled query-answer data. NUDGE-M and NUDGE-N offer two variants of the optimization; the repository describes evaluation on several text datasets and includes an example workflow.
Key Features
- Fine-tunes document embeddings without changing model parameters.
- Provides NUDGE-M and NUDGE-N through a shared interface.
- Uses pre-computed data and query embeddings with known relevant-record indexes.
- Accepts validation queries and answers as part of the fine-tuning workflow.
- Supports a memory-saving path for large corpora with records that are not positive answers to training or validation queries.
- Includes an example using NFcorpus and lists additional supported dataset names in the README.
Use Cases
- Retrieval engineers can adapt an existing embedding-based search system when they have labeled queries and relevant documents.
- RAG developers can test whether task-specific embedding adjustments improve document retrieval before passing results to a generation stage.
- Researchers can reproduce or extend experiments on non-parametric embedding fine-tuning, using the paper and the repository's experiment instructions.
- Teams working with large document collections can use the provided filtering optimization when fine-tuning all records at once strains available memory.
Getting Started
Install the package:
pip install nudge-ft
See the repository README for input formats, examples, dependencies, and experiment instructions.
Alternatives
- rerankers: rerankers improves retrieval results by reranking candidate documents at query time, while nudge adjusts precomputed document embeddings using labeled examples.
| Project | Language | License | Stars | Status |
|---|---|---|---|---|
| nudge | Python | MIT | 43 | Not checked yet |
| rerankers | Python | Apache-2.0 | 1.6k | Quiet |
Considerations
- NUDGE operates on embeddings; it does not embed documents or queries. You must provide embeddings and ground-truth answers for training and validation queries.
- Its optimization depends on labeled retrieval data, so it is not a plug-in replacement for embedding generation or a method for unlabeled data.
- The README warns that fine-tuning large corpora on a GPU can run out of memory. It describes a memory-saving option for certain records, but does not provide general hardware requirements.
- The reported example results are for NFcorpus and should not be assumed to transfer to other datasets or retrieval setups.
Found this useful?
Share it with someone who would like nudge.