Hello!
I am considering upgrade from DT3.
It would be nice if you could clarify these points for me because they are really important.
Do AI implementations in DT4 support the following features? First of all: Do you use RAG for all the notes and books that exist in my personal database? Do you update it with new notes and documents?
All what I ask below I already saw with my own eyes in some new note-taking apps. So it is already possible to implement.
Features:
Just tell the AI assistant to create a note with the dictated content and tag, and it will create the note with the dictated content and tag. Will it be supported in DT4 for iOS in voice mode?
Ask to find all notes with the mentioned content in plain english, and it will show all the notes containing that content.
Ask to improve the writing in a given note, and it will enhance it accordingly.
Ability to ask questions about a number of specific notes and/or PDF documents.
Ask to find notes with a particular image, and it will show me all the notes that contain that image.
DEVONthink 4 supports RAG but doesn’t use a vector database. E.g. its chat assistant can access selected documents and optionally change their properties/contents or perform e.g. database & web searches if enabled in the settings.
We don’t announce features of unreleased versions. But on the Mac you could create a voice note via the Sorter and optionally transcribe it. In DEVONthink To Go 3 it’s also possible to create voice notes and assign a tag.
Depends on content. Usually even DEVONthink’s search should be sufficient for this (and is much faster).
#3 is possible if enabled in the settings, otherwise via e.g. scripts (examples included). #4 is also possible (but the number is not unlimited), #5. isn’t.
But in the end this depends also highly on the used service, model and the type of document. E.g. the possibilities of local models are highly limited and the contents of PDF documents or proprietary third-party file formats can’t be modified.
Thank you very much for your answer. If possible, can you please briefly explain why you do not use vector database, as it is common solution for such cases nowadays?
You see, problem is in following:
With traditional search, you might end up retrieving:
Either too many loosely related results (e.g., full documents based on keyword match).
Or not enough semantically aligned content — it might miss the point. (I already experienced that several times)
Lastly, it consumes more context of the model on average, which means more money spend for query.
Do you have plans to implement vector database as part of Roadmap?
It’s an option for future releases. But so far DEVONthink 4’s chat assistant is optimized to work with selected items (and search results). This has its own benefits, e.g. privacy and frequently better results.
Thank you — it’s great to hear that this is being considered for future releases. From my perspective, it’s becoming increasingly clear that we are moving extremely rapidly toward a future where AI-based workflows will be the norm. In that context, adopting modern AI capabilities — such as vector-based semantic search — will be essential for staying relevant and competitive. I really hope DEVONthink continues to evolve in this direction.
Yes, and I truly appreciate that you’ve already integrated external AI capabilities. My point is that within the next couple of years, it’s very likely that users will come to expect seamless and deeply integrated AI features in all apps. Software that doesn’t offer intelligent, context-aware interactions — like semantic search, summarization, and smart agents — may quickly feel outdated, regardless of how strong its core features are. I believe DEVONthink is in a great position to lead in this space, and I’d love to see it evolve accordingly.
Version 4 integrates AI already in various ways and the interactive chat assistant is not necessarily the most important one. E.g. automation via batch processing or smart rules or scripting supports this too.
That’s definitely true—automation features like batch processing, smart rules, and scripting are indeed powerful and extremely valuable. My perspective is just that the real transformative potential of AI lies specifically in semantic understanding and context-aware interactions.
Current problem is like this. I already have this unpleasant limitations, especially second point.
Semantic Search: Finding Notes Without Exact Keywords
Current Situation:
You remember writing down something about “productivity strategies,” but you used different terminology in the actual note—perhaps something like “workflow optimization techniques.”
Traditional Search: You search for “productivity strategies” but don’t find the relevant note unless you remember the exact wording.
Semantic AI Search: You simply type “productivity strategies,” and DEVONthink retrieves your note mentioning “workflow optimization techniques,” understanding the meaning behind your query.
Context-Aware Assistant that Answers Directly from Your Notes
Current Situation:
You stored multiple notes and web archives related to a topic (e.g., investment strategies, coding best practices, or health advice). To find specific advice, you must manually read multiple notes.
Traditional Automation: You set up Smart Rules or tags, but ultimately you still have to read through each note to find exact answers.
Semantic AI Assistant: You simply ask a question like, “What are recommended strategies for diversifying a portfolio according to my notes?” and DEVONthink returns a concise, meaningful answer directly sourced from your own notes.
Of course. What I don’t understand? I wrote above only about what is already possible and what I used in some prototype note taking apps which was build with the implementation of vector databases. That’s why I am talking about it here. Because it is earth and heaven of difference in terms of relevance of search and queries.
I use Sonnet 3.7 in Devonthink.
Honestly, I was a bit confused by the words “semantic search” in your documentation. Because it is actually not semantic search — it just constructs queries to the database using exact words and does not cycle through all semantically close variants to find all the notes that are truly semantically close to the query (to at least imitate true semantic search based on vector database), like in the example above.
In DEVONthink even after enabling the database search in Settings > AI > Chat the search remains limited to the current selection in the item list or, if there’s none, to the selection in the sidebar. Again to increase privacy and to ensure that e.g. your financial or medical documents from another database aren’t accidentally used.
I think it would better be optional setting. Ideally to implement vector database with ability to choose what will go in it. Like add of external documentations in Cursor.
I don’t know where you’re reading about it, but the very few instances where “semantic search” are mentioned don’t make any wrong claims. If you think they do, then provide the page and excerpt.
Like add of external documentations in Cursor.
Cursor and DEVONthink are two wildly different applications.
About semantic search — it’s my fault that I wasn’t attentive enough to notice that it actually generates standard queries.
I used Cursor’s implementation of a vector database just as a convenient example. Your colleague already mentioned that you are considering a vector database in future upgrades, so it’s not impossible to implement and you’re even thinking about it. I just want to fully support this and emphasize that I believe it’s an essential feature of any contemporary knowledge database. It would be a pity if you decided to ignore it.