The risk is very duly noted.
If it were clear that Elephas only reads data and does not write to the database at all, are there any specific risks you could hypothesize about?
Does the act of reading data from the database alter any metadata or logs - if so does create any potential issues with syncing the database?
No specific risks to discuss off-hand though if it is modifying the documents in any way, you shouldn’t use it. We don’t develop it or have any concrete focus to try and make it “safe”. If we did, then why not AnythingLLM, GPTAll, Msty, or your Uncle Fred’s Newfangled AI too?
The stance is on principle: People shouldn’t be accessing documents inside the internals of a database, not from any app. If you want to use this app (or similar) for this purpose, I’d recommend you move the documents out to a safe local directory and index it.
If Elephas indexes your complete database, there’s most likely the risk that it might send undesired data (e.g. financial, personal or medical information) to a cloud service. Especially as the internal structure of DEVONthink’s databases does not match the visible group hierarchy, it’s probably impossible to control what’s going to be used or not.
Caveat Emptor - I’m not an expert in this area.
ObsidianCopilot (Copilot for Obsidian - The Ultimate AI Assistant for Your Second Brain) is a tool being developed to allow you to chat with your notes. Based on my limited understanding something similar could be done with DEVONthink - with a few massive caveats.
Steps 1 and 3, 4 - I know what is happening, step 2 is a guess.
FWIW Copilot alllows you to ask questions, etc of your vault. The developers call it NotebookLLM + your notes.
Limitations
I don’t want to tell @BLUEFROG and @cgrunenberg what to do.
And just to put things into perspective for any onlookers, many of our customers have one PDF that’s larger than their Obsidian vaults 
RAG on an Obsidian Vault, or even more easily in Drafts, would be child’s play compared to the massive volumes of data in peoples’ databases.
In that regard - I tried Elephas by copying some DT documents to a local folder and then adding them to Elephas. I compared its performance to some other AI tools which upload data to Claude.
Elephas performance is poor when querying a set of PDF documents in a “Superbrain.” It misses some important content entirely - and even worse it hallucinates often, something highly undesirable for a RAG system. And you cannot adjust the Temperature to reduce hallucinations.
My understanding is that Elephas is a hybrid model that uses local data indexing and then uploads the retrieved local data to the cloud when a query is made. The result in my view is that it is as fast as pure cloud AI but as poor quality as local AI.
The concept of their software and the Elephas UI are great but in my view it is a big fail with regard to actual usefulness. It gives me greater respect for AI software which is much more performant.
Actually I might be easier than you think.
If you just use an embeddings model as an aid in search, you’re not doing RAG.
In which case the model is only running on an expanded index. Since DT already maintains an index of the document text, the change in size will a factor of 2-3 times, not 10 times. So it is within the realm of possibility.
What is the weak point with Elephas and other hybrid models? Is it the local indexing and embedding? Whereas the more performant software use the larger (and thus more powerful) online models perform the indexing?
I suspect the weak point is the local indexing and embedding because I have compared the difference with the same LLM such as Claude 3.5 and the Anthropic web app or Typing Mind with Claude 3.5 give much better responses than Elephas with the same model.
You can use various models such as openai, claude, or any from openrouter. Some models allow temperature settings.
As for the embedding, you can also use the openai large embedding model, or select from alternative models.
I stick with the large v3 embedding model. The costs are negligible compared to the smaller models, but the performance difference is noticeable.
I have noticed that it can miss important information because, as far as I know, what happens is that the query is sent to the vector system which selects “excerpts” that “match” what may be required, and then sends the extracts to the LLM you select (openai, claude, openrouter) along with your query to provide the LLM context which the LLM uses to help answer the question.
I haven’t used Elephas with large pdfs since I mostly extract information from texts, but I can imagine that there is a learning curve as to how best use it with those sources, but for more self contained notes it is quite helpful.
Also, you can adjust the token content length, so for example I ask it to include up to 100k context which might be helpful in answering your question.
I use the superbrain in a one question / one answer fashion. I don’t think it’s good for a “conversation” type question and answer session.
Thanks @anonny
I suspect overall Elephas is best for shorter documents and when the goal is primarily to do a semantic search to locate documents.
Cloud-based AI is likely best when documents are larger and/or when the desired output is a detailed summary with source ilnks.
Hello everyone, I am Kamban, founder of Elephas. Thanks for mentioning Elephas. Elephas heavily relies on local computing and uses FAISS-based semantic search. Many of our users have reported good quality results. Combination of a good model and the system message (default message is tuned for this) has helped them achieve highly grounded results.
I am happy to review any quality issues if you can share more details. One feature we may add is the ability to rewrite queries before performing semantic search (as an option) to increase accuracy.
Claude-type services consume the whole document, while Elephas (RAG-based) uses chunks of data relevant to the search. Ideally suited for information discovery, and QA types of interaction.
With respect to privacy, you can run 100% offline with local models. Each Super Brain (a kind of workspace) can be configured to use a specific AI model. You need to use a local model for both indexing and chat sessions to ensure no data leaves your system. We have tested it with Llama 3.2.
Elephas does not write to the linked files. It uses read-only permission.
FYI: We have recently launched workflow automation on top of your data. This allows you to stitch a sequence of steps together to achieve a goal.
You can reach us at kamban at elephas.app or support at elephas.app for questions.
Welcome @kamban
I am curious about this. How does it use the search to chunk a document or is it chunking the document by default then running a quick search on the chunks to eliminate unrelated content?
Hi,
Yes, first it chunks the document as the first step. When searching, it scores each chunk and returns the top N results. You can control the number of results included for each chat session.
Hi @kamban
Are theses chunks created once when the document is indexed? Or are the chunks revised when there is a new query?
If the chunks are only created upon initial indexing then is it possible that some document hits will be missed?
I know the answer to this! 
Imagine it’s a database where the chunks are stored. When you input more data (index), then when you search the db you get the most relevant results even from the latest documents as long as they’ve been fully indexed.
You can see which are fully indexed in the folder.
Hi,
These chunks are updated when you add more data. So, at query time, all chunks will be searched against, including the new/updated ones.
Elephas has an auto-sync feature that can sync the changes from the source datasource.
A lot has changed in just 6 months or so.
Using ChatGPT-5 (with a $20/month Plus account) and a DropBox agent, I was able to execute conceptual searches of thousands of files across dozens of folders in my DropBox. I could ask questions like “Find any discussions of quantum theory that favor the Copenhagen interpretation” or “list my chess books that describe losing games by Garry Kasparov.” The searches took a while, but no West Coast servers erupted into flames or triggered droughts in Nevada (that I know of! ;))
That you know of.
When DT4 was announced I was quite happy that it would contain many AI features. And I had expected it would contain some RAG functions. RAG is no rocket science. Just extend the search index by a vector db index where besides plain content search now a similariy search will be an option. With good chunking the similarty search will give resonable results which should become part of the context for an LLM-AI chat about some documents of my DT database or parts of it.
Two week I asked Chat-GPT5 Pro how to setup a RAG based AI on top of a technical book in epub format I own. The answer looked sound and did contain just about 500 lines of code. For a PoC.
So I wonder why this RAG based AI chat is not an option for DT 4?
I own DT licenses since about 20 years. Private ones and also business. Privately it is DT 3 server. An not too cheap upgrade and from then on yearly fees make me wonder what useful AI features I will really gain form an upgrade.