Audrey Zhou

@VoyageAI

[turbopuffer @turbopuffer]

new: i8 vectors

f32: 4 bytes/dim
i8: 1 byte/dim

4x fewer bytes → 75% lower storage and query costs + faster queries

when embedded with a quantization-aware model (e.g. voyage-4-large) trained on i8 vectors, recall loss can be ~0!

docs: https://t.co/8brDT3O6X3

[Garry Tan @garrytan]

GBrain v0.27.1 now supports multi-modal image embeddings

[Audrey Zhou @laqqq_]

@ying11231 @radixark @lmsysorg @slime_framework @jxwuyi @xai Congratulations!!

[Voyage AI by MongoDB @VoyageAI]

To evaluate embeddings and retrieval, we need more benchmarks beyond MTEB that are less vulnerable to overfitting. That’s why RTEB was just beta-launched!

⚖️ Both open and held-out datasets to prevent overfitting to evaluation sets.
🌍 Realistic datasets from critical enterprise domains like law, healthcare, code, and finance.
🔎 Only focus on retrieval applications with relevant large-scale datasets.

Check out the blog and leaderboard on @huggingface and join the community in building a stronger, more reliable benchmark.

Blog: https://t.co/T6ZcROuVCg

[Frank Liu @frankzliu]

I feel like I've seen this kind of retrieval plot somewhere before 😂

[Voyage AI by MongoDB @VoyageAI]

📢 Meet voyage-3.5 and voyage-3.5-lite!
• flexible dim. and quantizations
• voyage-3.5 & 3.5-lite (int8, 2048 dim.) are 8% & 6% more accurate than OpenAI-v3-large, and 2.2x & 6.5x cheaper, resp. Also 83% less vectorDB cost!
• 3.5-lite ~ Cohere-v4 in quality, but 83% cheaper.

[Voyage AI by MongoDB @VoyageAI]

Voyage AI joined @MongoDB just 2 months ago, and we’re accelerating our mission of building the best embedding models for all developers!

Here’s what’s already in motion:
• 100x API scaling
• New, advanced models
• Built-in embeddings in Atlas Vector Search
• Direct API access in MongoDB Atlas

This is just the beginning. Follow along as we grow. https://t.co/8iFGoAp8f4

[Tengyu Ma @tengyuma]

RL + CoT works great for DeepSeek-R1 & o1, but: 

1️⃣ Linear-in-log scaling in train & test-time compute
2️⃣ Likely bounded by difficulty of training problems

Meet STP—a self-play algorithm that conjectures & proves indefinitely, scaling better! 🧠⚡🧵🧵

https://t.co/cZmBWOk0j9

[Hong Liu @HongLiu9903]

Tried to reproduce the COIR results. TLDR: SFR-Embedding-Code-2B_R is 26.5% worse than voyage-code-2 as oppposed to what is claimed in the paper.

[Audrey Zhou @laqqq_]

@doodlestein Super well written article

[Voyage AI by MongoDB @VoyageAI]

📢 Announcing voyage-code-3 embedding model!

1. more accurate: + 14% gain over OpenAI-v3-large
2. flexible dimension (Matryoshka): 256-2048
3. quantized embeddings: float, int8, binary
4. new Pareto frontier: (binary,256 dim.) is 6% better than OpenAI (float,3072 dim.) 🧵🧵

[Audrey Zhou @laqqq_]

@AlexFellowsFund @lilianweng @OpenAI Wow! So excited about Lilian exploring the venture world

[Audrey Zhou @laqqq_]

Evaluating the code retrieval quality of embedding models has been a common pain point.
Current community lacks high-quality benchmarking datasets. Builders suffer from noisy labels, a lack of deep algorithmic reasoning, and massive data contamination and overfitting problems.

[Voyage AI by MongoDB @VoyageAI]

Vector-based code retrieval is a critical building block in code assistants and agents. However, many people complained about the lack of diverse, high-quality evaluation datasets for it. We surveyed existing ones and proposed some methods to build better ones. 🧵🧵

[Voyage AI by MongoDB @VoyageAI]

Vector-based code retrieval is a critical building block in code assistants and agents. However, many people complained about the lack of diverse, high-quality evaluation datasets for it. We surveyed existing ones and proposed some methods to build better ones. 🧵🧵

[Audrey Zhou @laqqq_]

CLIP architecture will never be able to capture the whole semantics of multimodal input since it embeds texts and pics separately.
Unified embeddings for interleaved texts and images should be the way to go: https://t.co/9yrEIzPmxz

[Jina AI @JinaAI_]

Jina-CLIP-v2: a 0.9B multilingual multimodal embedding model that supports 89 languages, 512x512 image resolution, 8192 token-length, and Matryoshka representations down to 64-dim for both images and text. https://t.co/26xLxLKvBl With of course strong performance on retrieval & classification tasks. Like Jina-CLIP v1, the text encoder of Jina-CLIP v2 can function as a standalone dense retriever, giving performance comparable to jina-embeddings-v3, which is currently the best multilingual embedding model under 1B parameters.

[Audrey Zhou @laqqq_]

Voyage multimodal embedding models are powering better RAG. The previous RAG pipeline is basically half broken since companies will need to engineer a complex pipeline to turn the pdf/table/multimodal data into text in order for the information to be consumed.

[Voyage AI by MongoDB @VoyageAI]

📢 Announcing voyage-multimodal-3, our first multimodal embedding model!

It vectorizes interleaved text & images, capturing key visual features from screenshots of PDFs, slides, tables, figures, etc.

+19.63% accuracy gain on 3 multimodal retrieval tasks (20 datasets)! 🧵🧵

[Voyage AI by MongoDB @VoyageAI]

📢 Announcing voyage-multimodal-3, our first multimodal embedding model!

It vectorizes interleaved text & images, capturing key visual features from screenshots of PDFs, slides, tables, figures, etc.

+19.63% accuracy gain on 3 multimodal retrieval tasks (20 datasets)! 🧵🧵

[Voyage AI by MongoDB @VoyageAI]

For more details, check out our latest blog post: https://t.co/dCiibTnN0X

Start building with @VoyageAI today - the first 200M tokens are on us! We’d also love to support academic retrieval research and benchmarking. Please email us at [email protected] for more free tokens.

[Voyage AI by MongoDB @VoyageAI]

No need anymore for screen parsing models, layout analysis, or other complex text extraction pipelines. Simply take a screenshot of your document and/or prepend extra text (e.g., meta info), and vectorize the interleaved data! Start with our notebook:
https://t.co/QIbj9zO3wM

[Audrey Zhou @laqqq_]

Super exited!!

[Voyage AI by MongoDB @VoyageAI]

📢 Announcing voyage-multimodal-3, our first multimodal embedding model!

It vectorizes interleaved text & images, capturing key visual features from screenshots of PDFs, slides, tables, figures, etc.

+19.63% accuracy gain on 3 multimodal retrieval tasks (20 datasets)! 🧵🧵