We have released Ai Chat!
We are looking for feedback and first impressions. Please add any feedback on in the thread below.
This feedback will be considered by the product team for our first round of updates in the coming weeks.
Please note, this is not a place for submitting bugs. If you are encountering a bug, please file a report in app via 'submit a bug' in the bottom left settings menu in your Tana workspace.
8 Comments
When I use AI Chat on the left side without tagging any nodes, my assumption is that it can search the entire workplace or graph, rather than generating its own answers. It's crucial for AI Chat to scan through the entire dataset, which might be achieved through RAG or other technologies. Like the technology used in Notion AI, please consider adding tje capabilities . Thank you.
The only standout issue for me is that the chats don't feel part of the entire Tana ecosystem, as they are not themselves supertags. My hope is that the chats themselves become supertags, and then we get a supertag for #Assistants where we can create our own "custom GPTs" to apply to each chat. My dream workflow is to type out my query as a normal node, tag it #ai-chat and have it immediately give me the answer, with the option to quickly replace the text with the chat's answer (perhaps this is already possible, but I'm dumb). But I have to say that the UI is beautiful and I love the direction that you are going with it! It's awesome to have an AI that recognizes and responds in TIF.
Right now the default AI chat is saved to the root node's daily note, and would like to have the ability to choose to save to a certain workspace by default.
In tana, I'm very careful to separate my own content from AI generated content. I created an AI workspace specifically to store AI-generated content, so that I can refer to it in other workspaces without affecting the content that I originally wrote in my own.
In terms of AI renewal, it can only be renewed according to English, there seems to be no function option to change the default language, I hope to add a function to select a different language for renewal.
nice feature but to be honest quite useless right now because the ai chat does not have access to related content of the node. so the core function of tana (connect & relate & have access to it) is not part of the ai chat. and it cannot answer simple questions so i doubt that it can be really used to deal with complicated things. in your slack i wrote an example with screenshots. in my daily node i wrote "my brother send me an email today". then i used the ai chat, refecenced my daily note and asked: who wrote emails to me today? and the ai chat wasnt able to answer. other references also did not work. so it cannot get easier than this question.
Would be good to receive some type of guidance on how to best use this feature. For all my #contacts in Tana I have fields with details to each contact (like expertise, function, grade etc). Seems that the AI Chat does not receive that information as context. Even if I build a search with that information viewed as a table.
Feedback on AI Chat User Experience
I found that AI chat freezes due to an infinite loop in the child nodes under the search nodes. Specifically, the search node contains another search node, which then searches for the current AI chat node, leading to a deadlock and requiring a client restart.
In theory, In the current implementation logic， if the search conditions are sufficiently complex and precise, and the AI is provided with relevant content within a suitable scope, it should be able to answer questions effectively. However, the current process is not smooth. The main issues are concentrated in two areas:
Giving to the AI chat context by adding a search node leads to a higher threshold for using AIchat. This is because the threshold for writing complex search node statements is inherently high;
The results of searching nodes cannot be sorted by calculating the relevance like technologies like embedding and rerank, and only those with relevance of 0.85 or more can be provided as context. If all the search results are thrown to the AI, it is easy for the program to crash, and the output is hard to guarantee.
If the design goes in the direction it's going now, maybe it will be designed this way later? it might be necessary to design an agent that can generate its own search conditions. For example, if I ask, "What did I talk about with someone recently?", the process should automatically search for "communication-related" tags associated with that person and provide the search results as context to the AI.
This seems like it will get more and more complicated, unlike the popular RAG solution that allows you to ask questions directly to all the note content. This experience should be much better.
Perhaps the RAG (Retrieval-Augmented Generation) approach is unavoidable. I'm a loyal user of tana and I've been looking forward to the moment when tana supports RAG at every turn for the past year
Yeah, I second that. And with emphasize. I feel kind of tricked - and I know that was never the intention - but pushing and placing this AI feature in the spotlight where I've built - alone - API's that I can call from Tana that are delivering the same (or better) results is a letdown.
I hoped that tana will keep improving and focusing on the missing pieces in the very core, note taking and personal knowledge management, but now it's all AI as the rest of the world. And it's sadly another AI implementation that does only look good on paper I feel.
Seems to sell, or at least the hope is in that bucket.