UIBK-Embedding-v1

UIBK-Embedding-v1

Finds the passages that best match a query, using one model that combines token level and whole text matching. An English retrieval model from the University of Innsbruck for people building search.

Made by DataScience-UIBK on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a retrieval model for English text. The data science group at the University of Innsbruck trained it starting from ModernBERT-base. Search systems usually choose between a single vector per text and many token vectors per text. This model produces both from one encoder pass and keeps them in a single index, so you run one model instead of two.

Who it suits: Developers and researchers who build search or retrieval over English text, and who already use or want to try PyLate and late interaction retrieval.

Industry | Who uses it | How they use it | Example
Search engineering | Developers who build document search | Index documents with the model and rank them against queries | Searching a set of support articles with one index instead of two
Research | Information retrieval researchers | Compare it with other retrievers on benchmarks such as BEIR | Checking its average score against another ColBERT model
Knowledge tools | Teams that build question answering systems | Use it as the retrieval step before an answer is written | Fetching the passages a question refers to

Skip it if your text is not in English, or if you need a model that writes answers, since this one only ranks text.

Where to use it: The model loads as a PyLate ColBERT model and works with PLAID indexes as they are. The files download on Windows, macOS or Linux.

• PyLate: the library the model is made for
• PLAID indexes: the indexing method it works with
• MTEB: the benchmark suite the authors used to measure it
• Windows, macOS or Linux: where you can fetch the files with Python

Use case | Who | What they get back
Document retrieval | Search developers | A ranked list of matching passages
Benchmark comparison | Researchers | Scores to compare with other retrievers
Retrieval for question answering | Application teams | Candidate passages for a later step

How to set it up: You need Python and the PyLate library. The model files come from Hugging Face.

• Python with pip
• The PyLate library
• A small amount of disk space, about 596 MB for the F32 weights

1. Install the Hugging Face hub library.
2. Download the model files.
3. Load them as a PyLate ColBERT model and build your index.

pip install -U huggingface_hub
hf download DataScience-UIBK/uibk-embedding-v1

How to use it: Load the model in PyLate as you would any ColBERT model, then index your documents and score queries against them. Documents can run to 2048 tokens and queries to 256 tokens.

Tool | What it does
hf download | Fetches the model files to your computer
PyLate | Loads the model and scores queries against documents
PLAID | Indexes the vectors for fast search

Example: You index a set of English articles with the model, then search with a question. The ranking comes back from a plain token matching score over one list of vectors, which already includes the whole text score.

For newcomers and experts: If you are new to this, think of it as a way to rank passages by how well they answer a question. Start by indexing a few dozen short English documents and try one query.

If you know the field, each text becomes a set of vectors with 128 dimensions, as in any ColBERT model. Three extra vectors in that list carry the dense embedding, so plain MaxSim equals the token score plus a weighted dense score. The card says it was trained with knowledge distillation from a public cross encoder.

One encoder pass reads the text
Token vectors and dense vectors form one list
One PLAID index stores the list
Plain MaxSim ranks the results

Good to know: The authors report their own numbers.

Title: BEIR average nDCG@10, 15 datasets
Model | Average
UIBK-Embedding-v1 | 59.79
LateOn | 57.22

The card adds that this model is larger and reads longer documents than LateOn. LateOn climbs to 57.69 when the authors give it documents of 2048 tokens.

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the documents and queries you pass in
• Not for: non-English text, since the card lists English only

On Hugging Face, 13 November 2026

• Task: sentence similarity
• Library: PyLate
• Parameters: 149M
• Weights: about 596 MB in F32; about 75 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,612
• Likes: 11
• License: apache-2.0
• Languages: en
• Based on: answerdotai/ModernBERT-base
• Fine-tunes: 1
• Tags: pylate, safetensors, modernbert, colbert, multi-vector, sentence-transformers, sentence-similarity, feature-extraction, late-interaction, text-embeddings-inference

Links

• Model: https://huggingface.co/DataScience-UIBK/uibk-embedding-v1
• Files: https://huggingface.co/DataScience-UIBK/uibk-embedding-v1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download DataScience-UIBK/uibk-embedding-v1

Then follow the usage in its model card. Nothing in this kit runs by itself.