[HN Gopher] Kotaemon: An open-source RAG-based tool for chatting...
       ___________________________________________________________________
        
       Kotaemon: An open-source RAG-based tool for chatting with your
       documents
        
       Author : miles
       Score  : 159 points
       Date   : 2025-01-02 03:13 UTC (19 hours ago)
        
 (HTM) web link (github.com)
 (TXT) w3m dump (github.com)
        
       | jascha_eng wrote:
       | How well do these kind of pre built systems work? RAG in my
       | experience usually requires a decent amount of customization to
       | your input data for chunk formatting and other things to work
       | well, are these systems flexible enough?
       | 
       | What if you want to integrate it into an existing system, are
       | these for local use only? So everyone on e.g. a team has to set
       | it up themselves?
       | 
       | Kinda curious what the exact use case is I have seen a few of
       | these repos with 10k+ stars but don't really get what it's used
       | for.
        
         | bartread wrote:
         | I share some of your concerns with generalised/pre-built RAG
         | but on your local use only question, this is in the readme:
         | 
         | > Host your own document QA (RAG) web-UI: Support multi-user
         | login, organize your files in private/public collections,
         | collaborate and share your favorite chat with others.
         | 
         | Kotaemon does support customisation, so maybe that should allay
         | your concerns, but I do wonder how tricky it would be to
         | implement and then maintain.
        
         | Kiro wrote:
         | > RAG in my experience usually requires a decent amount of
         | customization
         | 
         | Like what? I'm curious because I just upload the documents to
         | OpenAI and make it available to the Assistant, and it seems to
         | work fine as a generic solution. Are they doing anything
         | magical?
        
           | infl8ed wrote:
           | >I just upload the documents to OpenAI and make it available
           | to the Assistant, and it seems to work fine
           | 
           | That is also my experience, OpenAI assistant attached docs do
           | seem to have a good amount of magic. Migrated over from an
           | admittedly basic/naive custom RAG solution and results are
           | similar/better but just have to work with a doc instead of
           | dealing with RAG. One thing I found is I have to add a strong
           | text to the prompt to force it to always check the doc, apart
           | from that works great.
        
           | diggan wrote:
           | > Are they doing anything magical?
           | 
           | It seems to me like they're doing a ton of magical stuff, but
           | it's really hard to know exactly without seeing the actual
           | source, since sadly their company name is a bit of misnomer.
           | Judging by the results of using it, they seem to be doing
           | some pre/post-processing to make it work with various of
           | formats and etc better.
        
           | dudus wrote:
           | When you do that you are just putting the whole book into
           | context for openAI to reason about. That works if the work of
           | smaller than the context.
           | 
           | For longer documents or for groups of documents you need a
           | kind of search to extract the most relevant passages to throw
           | in the context.
           | 
           | That is RAG. That search you do.
           | 
           | It's usually a semantic search using embedded data created by
           | a specialized model to create these embeds and a specific
           | algorithm to chunk the document into smaller pieces to derive
           | meaning from.
           | 
           | So you have multiple pieces involved into the job. A chunker,
           | an embbeder model, a vector database, etc.
        
       | behnamoh wrote:
       | not this again, we've already seen hundreds of such things...
        
         | exe34 wrote:
         | this is a very thoughtful comment, thank you for sharing it!
        
         | OtterBehemoth wrote:
         | I won't comment on here much, but why even post some of this
         | response? This is a nice piece of work.
        
       | antman wrote:
       | At this point there are multiple home RAG systems, pretty much
       | using the same components, so does anyone know how this compares
       | to others? I see that it imports GraphRAG which most other don't
        
         | jawns wrote:
         | I can't speak for the internals, but I've found it to be dead
         | simple to spin up locally and use, with decent results on the
         | docs I've tested it with.
         | 
         | That said, I think a lot of AI chat services have recognized
         | that document search is table stakes and are building this
         | functionality into their tools, so I don't know whether
         | Kotaemon as a standalone tool will be needed for much longer.
         | 
         | For example, my company was originally going to push out
         | Kotaemon for private document search, but we have now put that
         | on pause because we're exploring whether we can get the same
         | results through our primary AI chat service, without having to
         | point users to a separate tool.
        
           | sdesol wrote:
           | > whether Kotaemon as a standalone tool will be needed for
           | much longer.
           | 
           | I think a lot of AI startups will find themselves in this
           | situation. Searching and summarizing docs is a no-brainer for
           | OpenAI, Anthropic, etc. The only issue they have right now is
           | that their models might not be reliable enough due to the
           | non-deterministic nature of LLMs. In the long term, I believe
           | Google, Amazon and Microsoft will probably be the big winners
           | in this area since they can offer multiple models from major
           | providers to de-risk things.
           | 
           | Unless AI complements an existing solution that is unique
           | and/or is done well by existing businesses, it will be very
           | difficult to compete.
        
       | manishsharan wrote:
       | This looks great. Integrations with Graphrag framework is
       | helpful.
       | 
       | Are you able to monitor Token and Cost per user and per session?
       | 
       | Also, lots of times users have the same question worded
       | differently .Is there a cost effective way of answering them from
       | cache?
        
       ___________________________________________________________________
       (page generated 2025-01-02 23:01 UTC)