Story

tencent/EVIE-4.5B Visual Document Retrieval • 5B • Updated less than a minute ago • 1.97k • 34

---

license: apache-2.0 library_name: colpali-engine pipeline_tag: visual-document-retrieval base_model: tencent/EVIE-Preview-4.5B tags:

• colpali-engine
• qwen3_5
• vision-language
• colbert
• late-interaction
• multi-vector
• matryoshka
• vidore
• token-compression
• sentence-transformers datasets:
• vidore/vidore_benchmark
• vidore/vidore_benchmark_v2
• jinaai/jina-vdr

---

🏆 EVIE: The Most Accurate and Lightweight Visual Document Retriever

Evidence-Vector-Informed Embedding (EVIE)

High-Precision Late-Interaction Retrieval • Dynamic Prefix-MRL (64D–2048D) • Training-Free HAC Token Compression

🤗 EVIE-4.5B (Prefix-MRL & HAC)  •  🤗 EVIE-8B (Flagship Teacher)  •  🐙 GitHub: Tencent/EVIE

---

📢 Release Announcement: All model weights, training pipelines, token compression algorithms (HAC), and evaluation suites have been fully open-sourced in this repository. Full technical details, architectural ablations, and the formal research paper will be updated in an upcoming release.

---

🌟 Highlights

• Top-Tier Benchmark Performance: 66.75 on ViDoRe V3 for EVIE-8B and 66.02 for EVIE-4.5B with single-projection Prefix-MRL.
• ⚡ Prefix-MRL Elasticity: Single 2048D linear projection. Freely truncate at runtime into ${64, 128, 256, 512, 1024, 2048}$ dimensions without separate models.
• 📦 Ultra-Compact Index (HAC): Training-free Hierarchical Agglomerative Clustering compresses token counts from ~750 down to 32 vectors/page, slashing index storage to 3.81 GiB per million pages.
• 🌐 138 Multilingual Tasks Evaluated: Thoroughly evaluated across ViDoRe V1, V2, V3, and JinaVDR across 4 metric families (nDCG, Recall, MAP, MRR @1/5/10).
• 🔬 ARD Distillation Recipe: Anchor-preserving, capacity-aware relation distillation reproducing full student training from the 8B teacher.

---

🧠 Architecture & Technical Highlights

 Query Text  ────────► ColQwen3.5 (BiDir Attention) ────► Elastic Multi-Vectors (64D–2048D)
                                                                    │
                                                           MaxSim Matching
                                                                    │
 Doc Image   ────────► ColQwen3.5 (Vision Encoder)  ────► HAC Compression ──► 32 Vectors / Page

• Late-Interaction Multi-Vector Paradigm: Unlike dense single-vector retrieval that collapses high-resolution document pages into a single point, EVIE preserves fine-grained visual details (complex tables, layout structures, charts, and small typography) through token-level representations, scoring relevance via late-interaction MaxSim:

$$ S(Q, D) = \sum_{i=1}^{|Q|} \max_{j=1}^{|D|} (q_i \cdot d_j) $$

• Prefix-MRL (Single-Head Elastic Representation): EVIE-4.5B introduces single-projection Prefix-MRL. A single 2048D linear projection natively supports runtime truncation down to {64, 128, 256, 512, 1024, 2048} dimensions without maintaining multiple heads or separate checkpoints.
• ARD (Anchor-preserving Relation Distillation): The 4.5B student is distilled from the 8B teacher using token-relation topological geometry, hard-negative margin calibration, and anchor-preserving alignment, maintaining peak retrieval accuracy even under low-dimensional prefixes.
• HAC Token Compression (Hierarchical Agglomerative Clustering): A plug-and-play, training-free token reduction algorithm that aggregates visual patch tokens into 32 or 64 semantic centroids in joint feature-spatial space, reducing 1M-page index footprints to as little as 3.81 GiB.

---

🏆 ViDoRe V3 Leaderboard

Official ranking on the ViDoRe leaderboard (ViDoRe V3, Mean Task).

Rank	Model	Parameters	Mean (Task)
🥇 #1	tencent/EVIE-8B	8.4B	66.24
🥈 #2	tencent/EVIE-4.5B	4.5B	65.70
#3	webAI-ColVec1.1-8b	8.4B	64.95
#4	VultronRetrieverPrime-Qwen3.5-8B	8.4B	64.26
#5	webAI-ColVec1.1-4b	4.5B	63.90
#6	VultronRetrieverCore-Qwen3.5-4.5B	4.5B	63.57

---

📊 Comprehensive ViDoRe Leaderboard Comparison

Performance comparison across modern multi-vector late-interaction visual document retrievers on ViDoRe:

Rank	Model	Base Model	Param	Embed Dim	ViDoRe V1 (nDCG@5)	ViDoRe V2 (nDCG@5)	ViDoRe V3 (nDCG@10)
🥇	EVIE-8B	Qwen3.5-9B	8.41B	4096D	92.18	74.23	66.75
🥈	EVIE-4.5B	Qwen3.5-4B	4.61B	64–2048D Prefix-MRL	92.07	73.38	66.02
🥉	EVIE-Preview-4.5B	Qwen3.5-4B	4.54B	128D	91.73	70.87	65.36
4	webAI-ColVec1.1-8b	Qwen2.5-VL	8.40B	640D	91.30	65.82	65.32
5	VultronRetrieverPrime-8B	Qwen3.5-9B	8.40B	320D	92.08	68.18	64.26
6	webAI-ColVec1.1-4b	Qwen2.5-VL	4.54B	640D	90.49	63.60	63.90
7	VultronRetrieverCore-4.5B	Qwen3.5-4B	4.50B	320D	92.21	66.12	63.57
8	nemotron-colembed-vl-8b-v2	Nemotron-8B	8.80B	4096D	92.65	65.16	63.54
9	tomoro-colqwen3-embed-8b	Qwen2.5-VL	8.00B	320D	90.76	65.40	61.60
10	nemotron-colembed-vl-4b-v2	Nemotron-4B	4.80B	2560D	91.62	64.49	61.42
11	athrael-soju/colqwen3.5-4.5B-v3	Qwen3.5-4B	4.60B	128D	91.54	64.25	61.46
12	tomoro-colqwen3-embed-4b	Qwen2.5-VL	4.00B	320D	90.57	64.69	60.16
13	VultronRetrieverFlash-0.8B	Qwen3.5-0.8B	0.85B	320D	88.15	60.36	56.16

---

🔍 ViDoRe V3 Per-Domain Breakdown (nDCG@10)

Model	Avg	CompSci	Energy	Finance EN	Finance FR	HR	Industrial	Pharma	Physics
EVIE-8B	66.75	81.86	72.51	71.23	56.40	69.29	59.77	70.81	52.11
EVIE-4.5B	66.02	81.72	72.32	70.00	54.90	67.82	59.40	70.27	51.69
webAI-ColVec1.1-8b	65.32	80.08	70.12	71.90	54.87	68.55	57.65	67.88	51.50
nemotron-colembed-vl-8b-v2	63.54	79.30	69.82	67.29	51.54	66.32	56.03	67.19	50.84
VultronRetrieverPrime-8B	64.26	79.80	70.30	69.00	54.50	66.80	57.40	68.20	51.70
VultronRetrieverCore-4.5B	63.57	79.80	69.20	68.90	52.00	66.10	56.10	67.50	50.20
tomoro-colqwen3-embed-8b	61.60	75.35	68.41	65.08	49.10	63.98	54.41	66.36	50.13

---

🎯 Prefix-MRL Elastic Multi-Vector Head

EVIE-4.5B embeds document and query tokens with a single 2048D linear projection head trained via ARD. You can truncate the channel dimension on-the-fly without maintaining different models:

Full Projection (2048D)  [========================================================] 66.02
Prefix 1024D             [============================]                             65.94
Prefix 512D              [==============]                                           65.90
Prefix 256D              [=======]                                                  65.68
Prefix 128D              [===]                                                      65.27
Prefix 64D               [=]                                                        64.51

Dimension	Bytes / Vector	ViDoRe V1	ViDoRe V2	ViDoRe V3	JinaVDR	138-Task Avg4
64	128 B	92.16	73.18	64.51	81.00	77.71
128	256 B	92.28	73.37	65.27	81.83	78.18
256	512 B	92.20	73.83	65.68	82.09	78.45
512	1 KiB	92.38	74.53	65.90	82.27	78.77
1024	2 KiB	92.39	74.53	65.94	82.43	78.82
2048	4 KiB	92.53	74.91	66.02	82.48	78.98

---

🗜️ Token Compression (HAC)

Raw late-interaction representations keep all visual patch vectors (~750 vectors/page), requiring substantial storage. EVIE integrates Hierarchical Agglomerative Clustering (HAC) in a joint semantic-position space to cluster page tokens at indexing time:

$$ z = \text{L2}\left[(1-w)\text{L2}(v) + w p\right], \quad \mu_c = \text{L2}\left(\text{mean}_{i \in c} v_i\right) $$

Production Ready SKUs

SKU	Payload / Page	Index Size (1M Pages)	Vectors / Page	ViDoRe V1	ViDoRe V2	ViDoRe V3	Avg4
d64 K32	4 KiB	3.81 GiB	32	89.27	67.83	59.58	73.47
d64 K64	8 KiB	7.63 GiB	64	90.63	70.71	62.06	75.57
d128 K32	8 KiB	7.63 GiB	32	90.00	69.45	61.40	74.87

---

📋 Complete 138-Task Evaluation Matrix

Protocol paired-all-pages-dedup+process_queries+ndcg2r-20260827 ($\text{MVT} = 1024$, bidirectional attention):

Metric	ViDoRe V1	ViDoRe V2	ViDoRe V3	JinaVDR	4-Board Macro Avg
nDCG@1	88.35	72.76	61.16	74.26	74.13
nDCG@5	92.07	73.38	63.39	81.57	77.60
nDCG@10	92.53	74.91	66.02	82.48	78.98
Recall@1	88.35	36.73	30.14	74.26	57.37
Recall@5	94.95	65.59	58.06	87.39	76.50
Recall@10	96.36	76.56	69.67	90.17	83.19
MAP@1	88.35	73.18	65.03	74.26	75.20
MAP@5	91.09	66.52	55.02	79.61	73.06
MAP@10	91.29	66.25	55.50	79.99	73.25
MRR@1	88.35	73.18	65.03	74.26	75.20
MRR@5	91.09	81.66	74.76	79.61	81.78
MRR@10	91.29	82.11	75.43	79.99	82.20

---

⚡ Quick Start

ColPali Engine

Installation

git clone https://github.com/Tencent/EVIE.git
cd EVIE
pip install -r requirements.txt
export PYTHONPATH="$(pwd)/colpali${PYTHONPATH:+:$PYTHONPATH}"

Self-Contained Python Inference

import torch
from PIL import Image
from colpali_engine.models import ColQwen3_5, ColQwen3_5Processor
from colpali_engine.models.qwen3_5.colqwen3_5.modeling_colqwen3_5 import set_active_head

model_id = "tencent/EVIE-4.5B"

# 1. Load model with FlashAttention and bidirectional attention
model = ColQwen3_5.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    attn_implementation="flash_attention_2",
).eval()
model.enable_bidirectional_attention()

# 2. Select any Matryoshka head: 64, 128, 256, 512, 1024, or 2048
set_active_head(model, 128)

# 3. Process query and document image
processor = ColQwen3_5Processor.from_pretrained(model_id)
images = [Image.open("examples/demo/pages/q3_revenue.png").convert("RGB")]
queries = ["What is the total quarterly revenue?"]

image_batch = processor.process_images(images).to(model.device)
query_batch = processor.process_queries(queries).to(model.device)

# 4. Generate multi-vector representations and late-interaction score
with torch.inference_mode():
    image_embeddings = model(**image_batch)
    model.rope_deltas = None  # Reset RoPE deltas before text query forward
    query_embeddings = model(**query_batch)

scores = processor.score(query_embeddings, image_embeddings)
print("Late-interaction MaxSim Relevance Score:", scores)

Sentence Transformers

pip install "sentence-transformers[image] @ git+https://github.com/huggingface/sentence-transformers.git"

from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder("tencent/EVIE-4.5B")

queries = [
    "What is the variable represented on the y-axis of the graph?",
    "Total outlay is maximum in which year?",
]
documents = [
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc1.jpg",
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc2.jpg",
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc3.jpg",
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc4.jpg",
]

query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings[0].shape, document_embeddings[0].shape)
# torch.Size([23, 2048]) torch.Size([991, 2048])

scores = model.similarity(query_embeddings, document_embeddings)
print(scores)
# tensor([[17.2500,  7.6084,  6.7920,  4.8398],
#         [ 4.5312, 12.7812,  4.3530,  4.2368]])

Documents can be URLs, local image paths, or PIL.Image objects. The example returns the full 2048-dimensional token embeddings. Scores can vary slightly with dtype and attention backend. To use a smaller Prefix-MRL dimension, slice and renormalize:

from torch.nn.functional import normalize

dimension = 128
query_embeddings = [normalize(embedding[..., :dimension], dim=-1) for embedding in query_embeddings]
document_embeddings = [normalize(embedding[..., :dimension], dim=-1) for embedding in document_embeddings]
scores = model.similarity(query_embeddings, document_embeddings)

---

📂 Repository Layout

EVIE/
├── model.safetensors         # EVIE-4.5B weights (safetensors)
├── config.json               # Model configuration (Prefix-MRL, max 2048)
├── processor_config.json     # Multimodal processor config
├── infer.py                  # Standalone inference & scoring CLI
├── colpali/                  # ColQwen3.5 + Prefix-MRL + ARD
├── code/
│   ├── teacher/              # EVIE-8B training arms and soup merging
│   ├── student/              # EVIE-4.5B Prefix-MRL / ARD distillation
│   ├── shared/               # Data loaders, adapter merges, and 138-task eval harness
│   └── compress/             # Training-free HAC token compression pipeline
├── examples/demo/            # 8-page retrieval demo (run.sh)
└── env.sh.example            # Environment variables template

---

🔬 Training & Distillation Reproduction

Student training executes Prefix-MRL plus ARD: capacity-aware relation distillation and margin distillation against frozen EVIE-8B, with a Preview-anchor term on the 128D prefix:

cp env.sh.example env.sh
source env.sh

# Step 1: Train EVIE-8B teacher arms and merge
bash code/teacher/run.sh

# Step 2: Distill EVIE-4.5B student with ARD loss
export TEACHER_DIR=../Evie-8B
bash code/student/run.sh

# Step 3: Run comprehensive 138-task evaluation
MODEL_DIR=. RUN_NAME=evie-4.5b bash code/shared/eval_run.sh

---

🧩 Community Resources

Project	What it does	Link
ColNanoVDR	150M text-only query encoder distilled from EVIE-4.5B. Supports all Prefix-MRL widths and works with existing EVIE indexes, including HAC-compressed ones, so text queries can be encoded on CPU.	nanovdr/ColNanoVDR-Q-Ettin150M-EVIE45B-2048-ML

Community-contributed; not maintained or evaluated by the EVIE team.

---

📚 Citation

@misc{tencent2026evie,
  title        = {EVIE: High-Performance Multilingual Visual Document Retrieval with Matryoshka Embeddings and Token Compression},
  author       = {Wang, Zifei and Wen, Wei},
  year         = {2026},
  howpublished = {\url{https://github.com/Tencent/EVIE}},
  note         = {Corresponding author: Wei Wen <[email protected]>}
}

---

📄 License

This repository is licensed under the Apache-2.0 License.

---

👥 Authors

Name	Role	Affiliation	Contact
Zifei Wang	Author	Tencent IMA Product Center	[email protected]
Wei Wen	Author & Corresponding	Tencent Youtu Lab	[email protected]