After 6 months, can finally tell everyone what we have been working on. Looking forward to seeing what people will be able to build with this :)

Excited to share that Provence is accepted to #ICLR2025! Provence is a method for training an efficient & high-performing context pruner for #RAG, either standalone or combined with a reranker https://t.co/Q7TRYP0CPt w/ @thibault_formal @VNikoulina @sclincha @naverlabseurope

Today, we’re launching early access for North! Our all-in-one secure AI workspace platform combines LLMs, search, and agents into an intuitive interface that effortlessly integrates AI into your daily work to achieve peak productivity.

𝐋𝐚𝐮𝐧𝐜𝐡 𝐨𝐟 𝐂𝐨𝐡𝐞𝐫𝐞 𝐑𝐞𝐫𝐚𝐧𝐤 𝟑.𝟓 - 𝐁𝐨𝐨𝐬𝐭 𝐲𝐨𝐮𝐫 𝐒𝐞𝐚𝐫𝐜𝐡 🚀 What is new: - Large gains in multilingual retrieval 🇺🇳 - Reasoning Capabilities 🧮 - Strong gains Finance 🔢, eCommerce 👗, project management 📣 - New platforms: AWS Bedrock & Pinecone

Aya-Expanse, the strongest open weights multilingual LLM, was just released by @CohereForAI It beats Llama 70B multilingual, while being half the size and twice the speed.

Your search can see now. We're excited to release fully multimodal embeddings for folks to start building with!

Do not miss an application deadline for #ALPS2025 on October 15! https://t.co/j1NTFUxvNw ALPS is an Advanced Language Processing School, held in French Alps, with wonderful speakers, inspiring discussions around NLP, and outdoor activities such as skiing and hiking!

I will present our study on Multilingual Retrieval-augmented generation, tomorrow at #ACL2024NLP workshop on Knowledgeable LLMs, 16:00 (poster hall)! Come to discuss a multilingual extension of our BERGEN library: https://t.co/5uyOaRfEal https://t.co/FpBfP09x7P #NLProc

Looking for a team lead to join our search team at @cohere working with @Nils_Reimers and many other kind & smart people. Feel free to message me if you want to know more! here is the posting:

Announcing the private beta of our newest foundation embedding model, Cohere Compass: designed specifically for multi-aspect data like emails, invoices, CVs, and support tickets to offer superior enterprise search capabilities. Sign up to try it out! https://t.co/CO6OIrnpDg

0⃣ 𝐖𝐨𝐫𝐥𝐝 𝐅𝐢𝐫𝐬𝐭 𝐁𝐢𝐧𝐚𝐫𝐲 𝐕𝐞𝐜𝐭𝐨𝐫 𝐃𝐚𝐭𝐚𝐛𝐚𝐬𝐞 1⃣ Happy to annouce the world first 𝐁𝐢𝐧𝐚𝐫𝐲 𝐕𝐞𝐜𝐭𝐨𝐫 𝐃𝐚𝐭𝐚𝐛𝐚𝐬𝐞 (for educational purposes). 💰32x less memory 💰 🚀 40x faster search 🚀 Github: https://t.co/4jNgl0fROq

People are asking me what to expect from "Faster Learned Sparse Retrieval with Block-Max Pruning" with @ntonellotto and Torsten Suel 10x faster safe retrieval wrt. Maxscore and you can check yourself the approximate retrieval trade-offs on naver/splade-cocondenser-ensembledistil

Hope to see you later today in the poster session, and feel free to send a DM if you want to chat about this! Note that this was done while I was still at @naverlabseurope and it is a collaboration with @sclincha @ntonellotto and @HerveDejean.

We note that we promise this can be used conveniently but showcase results in PISA. This is actually also possible in ES. Actually the team at ES proposed a similar technique but without saturation or doc pruning ( https://t.co/uFsMKhS23W) concurrently with the ECIR deadline

In making SPLADE more efficient, one normally needs to either sacrifice convenience (still use the same tool, like ES) or OOD effectiveness. Not the case of Two-Step, in the 30 datasets we tested, 22 did not saw statistical drop in effectiveness, while being as efficient as BM25

Today I'm presenting Two-Step SPLADE at #ECIR2024 findings! We show how to run SPLADE with BM25-like latency and SPLADE-v3 effectiveness TLDR: Prune vectors, retrieve(top-k), rescore with unpruned ones. Paper: https://t.co/rd9ezd5IeF Github: https://t.co/8qVgdBX2al

🚀 𝐂𝐨𝐡𝐞𝐫𝐞 𝐄𝐦𝐛𝐞𝐝 𝐕𝟑 - 𝐢𝐧𝐭𝟖 & 𝐛𝐢𝐧𝐚𝐫𝐲 𝐒𝐮𝐩𝐩𝐨𝐫𝐭🚀 I'm excited to launch our native support for int8 & binary embeddings for Cohere Embed V3. They slash your vector DB cost 4x - 32x while keeping 95% - 100% of the search quality. https://t.co/uJBg6nyPvf

Just as splade-v3 comes out ( https://t.co/j1CeQIO4iL), splade++ achieves 1M monthly downloads again, pretty cool seeing this happen again! Always curious to know what people are doing it with and congrats to the @naverlabseurope team @thibault_formal @sclincha HerveDejean

🇺🇳𝟐𝟓𝟎𝐌 𝐖𝐢𝐤𝐢𝐩𝐞𝐝𝐢𝐚 𝐄𝐦𝐛𝐞𝐝𝐝𝐢𝐧𝐠𝐬 𝐢𝐧 𝟑𝟎𝟎+ 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞𝐬 🇺🇳 What could you build if your RAG has access to Wikipedia in all 300+ languages? Available for anyone to use, using our state-of-the-art multilingual embedding model: https://t.co/PCaAyboxsq

Welcome @CohereForAI Command-R! The top trending among over 500k open-access models! 🚀 https://t.co/F4ee7TAulJ