EigenLI's authors make the comparison useful by reporting the exception alongside the wins: spectral compression beats clustering on ColBERTv2 and AnswerAI-ColBERT-small at k≤32, while GTE-ModernColBERT favors clustering at k=32.
EigenLI: Spectral Approximations to Late Interaction
Microsoft presents a training-free spectral method that compresses ColBERT-style representations, outperforming clustering pooling and MUVERA single-vector baselines.
📝 arxiv.org/abs/2609.07561
Microsoft presents a training-free spectral method that compresses ColBERT-style representations, outperforming clustering pooling and MUVERA single-vector baselines.
📝 arxiv.org/abs/2609.07561
EigenLI: Spectral Approximations to Late Interaction
Late-interaction models such as ColBERT achieve strong effectiveness by representing each document with many token-level vectors, but this expressivity leads to large indexing cost, storage footprints...
arxiv.org
September 10, 2026 at 6:51 PM
EigenLI's authors make the comparison useful by reporting the exception alongside the wins: spectral compression beats clustering on ColBERTv2 and AnswerAI-ColBERT-small at k≤32, while GTE-ModernColBERT favors clustering at k=32.
And it works.
HippoRAG beat top systems like ColBERTv2 and IRCoT in 3 multi-hop QA benchmarks.
It retrieved answers:
✅ 6–13x faster
✅ 10–30x cheaper
✅ More accurately
It skips iterative queries and finds the answer in one shot.
6/9
HippoRAG beat top systems like ColBERTv2 and IRCoT in 3 multi-hop QA benchmarks.
It retrieved answers:
✅ 6–13x faster
✅ 10–30x cheaper
✅ More accurately
It skips iterative queries and finds the answer in one shot.
6/9
May 26, 2025 at 4:04 PM
And it works.
HippoRAG beat top systems like ColBERTv2 and IRCoT in 3 multi-hop QA benchmarks.
It retrieved answers:
✅ 6–13x faster
✅ 10–30x cheaper
✅ More accurately
It skips iterative queries and finds the answer in one shot.
6/9
HippoRAG beat top systems like ColBERTv2 and IRCoT in 3 multi-hop QA benchmarks.
It retrieved answers:
✅ 6–13x faster
✅ 10–30x cheaper
✅ More accurately
It skips iterative queries and finds the answer in one shot.
6/9
3) "Multi-vector" makes it sound like these approaches win because they store "more stuff".
But that's not true: if you look at how aggressive ColBERTv2 representations are compressed, it's often ~20 bytes per vector (like 5 floats), which can be smaller than popular uncompressed single vectors!
But that's not true: if you look at how aggressive ColBERTv2 representations are compressed, it's often ~20 bytes per vector (like 5 floats), which can be smaller than popular uncompressed single vectors!
February 26, 2025 at 7:14 PM
3) "Multi-vector" makes it sound like these approaches win because they store "more stuff".
But that's not true: if you look at how aggressive ColBERTv2 representations are compressed, it's often ~20 bytes per vector (like 5 floats), which can be smaller than popular uncompressed single vectors!
But that's not true: if you look at how aggressive ColBERTv2 representations are compressed, it's often ~20 bytes per vector (like 5 floats), which can be smaller than popular uncompressed single vectors!
Yes, ColBERTv2 via PLAID. (Even ColBERTv1, properly configured, had that property.)
The square root is the number of cluster centroids used for pruning almost all documents before scoring. (It's described in the ColBERTv2 paper. See the PLAID paper, though, for the optimized process.)
The square root is the number of cluster centroids used for pruning almost all documents before scoring. (It's described in the ColBERTv2 paper. See the PLAID paper, though, for the optimized process.)
January 4, 2025 at 6:59 PM
Yes, ColBERTv2 via PLAID. (Even ColBERTv1, properly configured, had that property.)
The square root is the number of cluster centroids used for pruning almost all documents before scoring. (It's described in the ColBERTv2 paper. See the PLAID paper, though, for the optimized process.)
The square root is the number of cluster centroids used for pruning almost all documents before scoring. (It's described in the ColBERTv2 paper. See the PLAID paper, though, for the optimized process.)
Finding 1: Agent-issued queries follow a web-search style (keywords, phrases, and quoted exact matches). BM25 (with suitable parameters) and multi-vector ColBERTv2 perform best, followed by learned sparse SPLADE-v3, while single-vector Qwen3-Embed-8B lags behind others.
April 9, 2026 at 4:44 PM
Finding 1: Agent-issued queries follow a web-search style (keywords, phrases, and quoted exact matches). BM25 (with suitable parameters) and multi-vector ColBERTv2 perform best, followed by learned sparse SPLADE-v3, while single-vector Qwen3-Embed-8B lags behind others.
A two-stage retriever using ModernBERT and ColBERTv2 lifted Recall@3 by up to 4.2 pts and reached 0.4448 accuracy on the MIRAGE benchmark, slightly beating MedCPT’s 0.4436. https://getnews.me/modernbert-colbert-enhance-biomedical-retrieval-augmented-generation/ #modernbert #colbertv2 #biomedicalai
October 8, 2025 at 4:26 AM
A two-stage retriever using ModernBERT and ColBERTv2 lifted Recall@3 by up to 4.2 pts and reached 0.4448 accuracy on the MIRAGE benchmark, slightly beating MedCPT’s 0.4436. https://getnews.me/modernbert-colbert-enhance-biomedical-retrieval-augmented-generation/ #modernbert #colbertv2 #biomedicalai
1️⃣ Set Up
We start by setting up the LM (Mistral-7B served using @Ollama & retriever model, RM which is ColBERTv2)
Check this out👇
We start by setting up the LM (Mistral-7B served using @Ollama & retriever model, RM which is ColBERTv2)
Check this out👇
December 12, 2024 at 12:30 PM
1️⃣ Set Up
We start by setting up the LM (Mistral-7B served using @Ollama & retriever model, RM which is ColBERTv2)
Check this out👇
We start by setting up the LM (Mistral-7B served using @Ollama & retriever model, RM which is ColBERTv2)
Check this out👇
🔎 LI papers:
📄 ColBERTv2 (NAACL 2022), 6-10x compressed token vectors, ‘denoised supervision’ as well as standard contrastive loss arxiv.org/abs/2112.01488
📄 SPLATE, sparse late interaction retrieval (SIGIR 2024) arxiv.org/abs/2404.13950
📄 MUVERA (2024) arxiv.org/abs/2405.19504
📄 ColBERTv2 (NAACL 2022), 6-10x compressed token vectors, ‘denoised supervision’ as well as standard contrastive loss arxiv.org/abs/2112.01488
📄 SPLATE, sparse late interaction retrieval (SIGIR 2024) arxiv.org/abs/2404.13950
📄 MUVERA (2024) arxiv.org/abs/2405.19504
ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction
Neural information retrieval (IR) has greatly advanced search and other knowledge-intensive language tasks. While many neural IR methods encode queries and documents into single-vector representations...
arxiv.org
December 26, 2025 at 7:27 PM
🔎 LI papers:
📄 ColBERTv2 (NAACL 2022), 6-10x compressed token vectors, ‘denoised supervision’ as well as standard contrastive loss arxiv.org/abs/2112.01488
📄 SPLATE, sparse late interaction retrieval (SIGIR 2024) arxiv.org/abs/2404.13950
📄 MUVERA (2024) arxiv.org/abs/2405.19504
📄 ColBERTv2 (NAACL 2022), 6-10x compressed token vectors, ‘denoised supervision’ as well as standard contrastive loss arxiv.org/abs/2112.01488
📄 SPLATE, sparse late interaction retrieval (SIGIR 2024) arxiv.org/abs/2404.13950
📄 MUVERA (2024) arxiv.org/abs/2405.19504
This innovative method uncovers origin of place names using retrieval augmented generation. By leveraging advanced language models like ColBERTv2 and Llama2, this research reveals intricate histories while addressing representation gaps in community narratives. https://arxiv.org/abs/2509.01030
Identifying Origins of Place Names via Retrieval Augmented Generation
ArXiv link for Identifying Origins of Place Names via Retrieval Augmented Generation
arxiv.org
September 4, 2025 at 6:02 AM
This innovative method uncovers origin of place names using retrieval augmented generation. By leveraging advanced language models like ColBERTv2 and Llama2, this research reveals intricate histories while addressing representation gaps in community narratives. https://arxiv.org/abs/2509.01030
Published a blog post version of my PLAID scoring pipeline code walkthrough with explanatory/narrative text!
vishalbakshi.github.io/blog/posts/2...
vishalbakshi.github.io/blog/posts/2...
December 24, 2024 at 5:25 PM
Published a blog post version of my PLAID scoring pipeline code walkthrough with explanatory/narrative text!
vishalbakshi.github.io/blog/posts/2...
vishalbakshi.github.io/blog/posts/2...
(2/2)
Colab: colab.research.google.com/drive/1pbF8o...
My initial exploratory notebook, very verbose and meandering, but with an additional deep dive into StridedTensors and the IVF: colab.research.google.com/drive/1zOo0C...
PLAID ColBERTv2 Paper: arxiv.org/abs/2205.09707
Colab: colab.research.google.com/drive/1pbF8o...
My initial exploratory notebook, very verbose and meandering, but with an additional deep dive into StridedTensors and the IVF: colab.research.google.com/drive/1zOo0C...
PLAID ColBERTv2 Paper: arxiv.org/abs/2205.09707
December 21, 2024 at 4:53 PM
(2/2)
Colab: colab.research.google.com/drive/1pbF8o...
My initial exploratory notebook, very verbose and meandering, but with an additional deep dive into StridedTensors and the IVF: colab.research.google.com/drive/1zOo0C...
PLAID ColBERTv2 Paper: arxiv.org/abs/2205.09707
Colab: colab.research.google.com/drive/1pbF8o...
My initial exploratory notebook, very verbose and meandering, but with an additional deep dive into StridedTensors and the IVF: colab.research.google.com/drive/1zOo0C...
PLAID ColBERTv2 Paper: arxiv.org/abs/2205.09707
Very excited to share my step-by-step walkthrough of the colbert repo code needed to recreate the 4-stage PLAID scoring pipeline to match RAGatouille results! Amazing to see centroids, scores, passage token encodings & passage IDs come to life!!
youtu.be/XRPP5LHHk0o
(1/2)
youtu.be/XRPP5LHHk0o
(1/2)
Recreating the PLAID ColBERTv2 Scoring Pipeline: From Research Code to RAGatouille
YouTube video by vishal
youtu.be
December 21, 2024 at 4:53 PM
Very excited to share my step-by-step walkthrough of the colbert repo code needed to recreate the 4-stage PLAID scoring pipeline to match RAGatouille results! Amazing to see centroids, scores, passage token encodings & passage IDs come to life!!
youtu.be/XRPP5LHHk0o
(1/2)
youtu.be/XRPP5LHHk0o
(1/2)
Reading ColBERTv2 paper helped me better understand vanilla ColBERT and now reading PLAID is helping me better understand ColBERTv2: partly bc exposure, partly bc imo the authors describe previous concepts MUCH more clearly in the next paper. I think it's bc maybe they understand it better as well?
November 25, 2024 at 11:22 PM
Reading ColBERTv2 paper helped me better understand vanilla ColBERT and now reading PLAID is helping me better understand ColBERTv2: partly bc exposure, partly bc imo the authors describe previous concepts MUCH more clearly in the next paper. I think it's bc maybe they understand it better as well?
Fantastic (small) Retrievers and How to Train Them: mxbai-edge-colbert-v0 Tech Report
Mixedbread AI introduces small but powerful ColBERT models (17M and 32M parameters) that outperform ColBERTv2 on short-text benchmarks.
📝 arxiv.org/abs/2510.14880
Mixedbread AI introduces small but powerful ColBERT models (17M and 32M parameters) that outperform ColBERTv2 on short-text benchmarks.
📝 arxiv.org/abs/2510.14880
Fantastic (small) Retrievers and How to Train Them: mxbai-edge-colbert-v0 Tech Report
In this work, we introduce mxbai-edge-colbert-v0 models, at two different parameter counts: 17M and 32M. As part of our research, we conduct numerous experiments to improve retrieval and late-interact...
arxiv.org
October 17, 2025 at 5:03 AM
Fantastic (small) Retrievers and How to Train Them: mxbai-edge-colbert-v0 Tech Report
Mixedbread AI introduces small but powerful ColBERT models (17M and 32M parameters) that outperform ColBERTv2 on short-text benchmarks.
📝 arxiv.org/abs/2510.14880
Mixedbread AI introduces small but powerful ColBERT models (17M and 32M parameters) that outperform ColBERTv2 on short-text benchmarks.
📝 arxiv.org/abs/2510.14880
ModernBERT + ColBERT: Enhancing Biomedical RAG Through an Advanced Re-Ranking Retriever
Combines ModernBERT bidirectional encoder with ColBERTv2 late-interaction re-ranker for biomedical retrieval.
📝 arxiv.org/abs/2510.04757
👨🏽💻 anonymous.4open.science/r/biorag-MC-...
Combines ModernBERT bidirectional encoder with ColBERTv2 late-interaction re-ranker for biomedical retrieval.
📝 arxiv.org/abs/2510.04757
👨🏽💻 anonymous.4open.science/r/biorag-MC-...
ModernBERT + ColBERT: Enhancing biomedical RAG through an advanced re-ranking retriever
Retrieval-Augmented Generation (RAG) is a powerful technique for enriching Large Language Models (LLMs) with external knowledge, allowing for factually grounded responses, a critical requirement in hi...
arxiv.org
October 7, 2025 at 4:23 AM
ModernBERT + ColBERT: Enhancing Biomedical RAG Through an Advanced Re-Ranking Retriever
Combines ModernBERT bidirectional encoder with ColBERTv2 late-interaction re-ranker for biomedical retrieval.
📝 arxiv.org/abs/2510.04757
👨🏽💻 anonymous.4open.science/r/biorag-MC-...
Combines ModernBERT bidirectional encoder with ColBERTv2 late-interaction re-ranker for biomedical retrieval.
📝 arxiv.org/abs/2510.04757
👨🏽💻 anonymous.4open.science/r/biorag-MC-...
DPR's 110M parameters fit neatly on Huawei's Ascend — small enough for CPU inference, but Contriever, ColBERTv2 and E5 already beat it on dense retrieval benchmarks. Hard to see rebuilding a RAG pipeline around this in 2026 when E5-mistral outperforms with one encoder swap.
Unlocking Dense Retrieval: Facebook's DPR Model Lands on Modelers.cn
aichina.news
July 17, 2026 at 11:50 AM
DPR's 110M parameters fit neatly on Huawei's Ascend — small enough for CPU inference, but Contriever, ColBERTv2 and E5 already beat it on dense retrieval benchmarks. Hard to see rebuilding a RAG pipeline around this in 2026 when E5-mistral outperforms with one encoder swap.
The MIT licence and 4-layer CamemBERT make this ColBERTv2 variant surprisingly viable for French RAG on Ascend. Yet, missing native inference benchmarks leaves latency gains unverified. Hard to see this beating standard bi-encoders at scale until the team posts real-world Ascend throughput data.
High-Performance French Ranking at Scale: Exploring the New ColBERTv2 CamemBERT for Ascend NPUs
aichina.news
April 30, 2026 at 8:57 AM
The MIT licence and 4-layer CamemBERT make this ColBERTv2 variant surprisingly viable for French RAG on Ascend. Yet, missing native inference benchmarks leaves latency gains unverified. Hard to see this beating standard bi-encoders at scale until the team posts real-world Ascend throughput data.
Models: huggingface.co/mixedbread-a... and huggingface.co/mixedbread-a...
Report: www.mixedbread.com/papers/small...
Blog: www.mixedbread.com/blog/edge-v0
Report: www.mixedbread.com/papers/small...
Blog: www.mixedbread.com/blog/edge-v0
Fantastic (small) Retrievers and How to Train Them: mxbai-edge-colbert-v0
Introducing our new family of extremely efficient ColBERT models, to serve as backbones for modern late interaction research while outperforming ColBERTv2 with just 17 million parameters.
www.mixedbread.com
October 17, 2025 at 3:59 AM
mxbai-edge-colbert-v0: tiny long context multi-vector embedding models
This report is huge, it gives us:
- Apache 2 17M(!!) and 32M models
- tops LongEmbed benchmark
- reproducible(!!) training pipelines
- extensive ablations to understand ColBERT models
www.mixedbread.com/blog/edge-v0
This report is huge, it gives us:
- Apache 2 17M(!!) and 32M models
- tops LongEmbed benchmark
- reproducible(!!) training pipelines
- extensive ablations to understand ColBERT models
www.mixedbread.com/blog/edge-v0
Fantastic (small) Retrievers and How to Train Them: mxbai-edge-colbert-v0
Introducing our new family of extremely efficient ColBERT models, to serve as backbones for modern late interaction research while outperforming ColBERTv2 with just 17 million parameters.
www.mixedbread.com
October 17, 2025 at 10:47 AM
mxbai-edge-colbert-v0: tiny long context multi-vector embedding models
This report is huge, it gives us:
- Apache 2 17M(!!) and 32M models
- tops LongEmbed benchmark
- reproducible(!!) training pipelines
- extensive ablations to understand ColBERT models
www.mixedbread.com/blog/edge-v0
This report is huge, it gives us:
- Apache 2 17M(!!) and 32M models
- tops LongEmbed benchmark
- reproducible(!!) training pipelines
- extensive ablations to understand ColBERT models
www.mixedbread.com/blog/edge-v0
paper links for the lazy:
ColBERT: arxiv.org/abs/2004.12832
ColBERTv2: arxiv.org/abs/2112.01488
PLAID: arxiv.org/abs/2205.09707
ColBERT: arxiv.org/abs/2004.12832
ColBERTv2: arxiv.org/abs/2112.01488
PLAID: arxiv.org/abs/2205.09707
January 4, 2025 at 7:03 PM
paper links for the lazy:
ColBERT: arxiv.org/abs/2004.12832
ColBERTv2: arxiv.org/abs/2112.01488
PLAID: arxiv.org/abs/2205.09707
ColBERT: arxiv.org/abs/2004.12832
ColBERTv2: arxiv.org/abs/2112.01488
PLAID: arxiv.org/abs/2205.09707
Is this ColBERTv2 or something later?
January 4, 2025 at 6:57 PM
Is this ColBERTv2 or something later?
ModernBERT + ColBERT: Enhancing biomedical RAG through an advanced re-ranking retriever [new]
Enhances biomedical RAG by combining ModernBERT for initial retrieval with ColBERTv2 for refined re-ranking to improve accuracy.
Enhances biomedical RAG by combining ModernBERT for initial retrieval with ColBERTv2 for refined re-ranking to improve accuracy.
October 7, 2025 at 3:35 AM
ModernBERT + ColBERT: Enhancing biomedical RAG through an advanced re-ranking retriever [new]
Enhances biomedical RAG by combining ModernBERT for initial retrieval with ColBERTv2 for refined re-ranking to improve accuracy.
Enhances biomedical RAG by combining ModernBERT for initial retrieval with ColBERTv2 for refined re-ranking to improve accuracy.
Use Neo4j for traversal constraints and structured context pruning, and ColBERTv2 for dense semantic re-ranking within the pruned subgraph.
August 31, 2025 at 7:36 PM
Use Neo4j for traversal constraints and structured context pruning, and ColBERTv2 for dense semantic re-ranking within the pruned subgraph.