Hi again everyone. I am excited to test the new auto OCR feature and the DeepSeek R1. I just noticed that Deepseek v3 and DeepSeek R1 were added before the weekend. Though they don't seem to be working yet. I'm especially excited for the thinking models like R1 since I tested some of them I was wondering if Dust will include some and how I can integrate some assistants using those thinking models. I hope DeepSeek R1 can be fixed and easy to integrate in the same thread with other models since I noticed that when using assistants from different models especially when some use tools there are very often errors when switching between them in the same thread (especially using Gemini after using claude). Any way If any one have tips for this issue using the system prompt please share them. And Dust team please let me know when will R1 be fully integrated.
They have not been enabled yet for all workspaces because at the moment they use directly Deepseek API (so data in China), we're currently working on having them available through Together AI with data in US, it should be ready soon. 🙏🏻 I could activate it for you with Deepseek API right away but there's a chance we remove support in the coming days/weeks because we're cooking something better on how to interact with reasoning models. So if you spend time building assistants based on those models they might not work anymore after we shipped this project.
Thanks Daphné if they can use tools with the China API I would love to have them enabled to test out the performance without diving into building in-depth assistants while waiting for the real thing. Otherwise I can still just use their free chatbot.
I think I got a hint of what you're cooking at dust when I saw the reasoning checkbox in tools. And I'm loving the idea already. We can still use our favorite assistants and give them superpowers when needing more reasoning heavy tasks, that's awesome. I'm curious to see how you will handle the mixing of reasoning and other tool uses and how we would be able to see the thinking process for debugging or other purposes. The context window is also something to think about, since reasoning models allocate a large portion of their output tokens for the reasoning process and if the reasoning tokens are fed back into the context window, it could more than half the available useful context window, so I'm suspecting the solution you are coming up with for reasoning as a tool might be, in part, for that specific limitation. Anyway I wish you really good luck and I'm eager to try whatever solution you are cooking.
Oh it's already around the corner. Though I don't know if it's an early announcement or something got wrong on launch cause the little reasoning box disappeared for me (even though it was showing yesterday). The R1 model is finally working properly after the hickups yesterdat and today for what appears to be ddos attacks on deepseek's servers. But the global model doesn't appear to have access to tools other than the visualization tool. I'm not being impatient (or am I 🙃), Just wanted to give feedback in case something isn't going as expected for the deployment.
Can you check again? 🙏🏻
Thank you again for your help Daphné
Hi Daphné I would love to have access to R1 and the reasoning tool as well on my workspace. Please activate the new feature flags. Really looking forward to test it. 🙂 Thanks in advance 🥳
Hi Luca Tjaden, we're rolling out progressively and I see your workspace is in a pilot plan so I've reached out to Thibaut internally to discuss it with you! 🙏🏻
Hey Daphné 👋 I had access to Deepseek but can't see it anymore, was it a rollout mistake on my account? 😿 Happy to see new reasoning models as it's a game changer
Hi Damien Mesgard I can't find an active workspace using your Slack email address, can you DM me your workspace id (you have it on the url when you're logged in)?
Done Damien Mesgard! ✅
