RDFox?

[Michael DeBellis]

I have a potential client who is interested in a new RDF triplestore called RDFox. I haven't had time to look at it in detail and would appreciate some feedback if anyone has used or done a serious evaluation of it. All I know now is that it stores all its data in RAM rather than on disk. So I would think that makes it very fast but also brings some challenges regarding scale and fault tolerance. 

Michael

https://www.michaeldebellis.com/blog

[David Poole]

If they just want to query an in-memory triple store, a very efficient triplestore can be implemented in 40 lines of Python. See Section 16.1 in knowledgeGraph.py of aipython.org <http://aipython.org/>.

It has no bells and whistles. But it depends on what they want.

It's open-source under a creative commons licence.

David

> On Jul 22, 2026, at 3:19 PM, Michael DeBellis <mdebe...@gmail.com> wrote:
>
> [CAUTION: Non-UBC Email]I have a potential client who is interested in a new RDF triplestore called RDFox. I haven't had time to look at it in detail and would appreciate some feedback if anyone has used or done a serious evaluation of it. All I know now is that it stores all its data in RAM rather than on disk. So I would think that makes it very fast but also brings some challenges regarding scale and fault tolerance.
>
> Michael
> https://www.michaeldebellis.com/blog
>
> --
> All contributions to this forum are covered by an open-source license.
> For information about the wiki, the license, and how to subscribe or
> unsubscribe to the forum, see http://ontologforum.org/info
> ---
> You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
> To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/e242fe79-068d-4fd0-9cf5-5c82ef168488n%40googlegroups.com.

[Michael DeBellis]

Interesting. Thanks, I'll add it to my list of OWL/RDF tools.  I don't think  aipython is appropriate for this project. But it is something I might find useful for other projects. Does it handle concurrent users and ACID transactions? Actually, if it does with 40 lines of Python that would be amazing! 

Michael

You received this message because you are subscribed to a topic in the Google Groups "ontolog-forum" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/ontolog-forum/U_Yku77l8lo/unsubscribe.
To unsubscribe from this group and all its topics, send an email to ontolog-foru...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/676204AF-FAE0-479E-911D-F696E8BECBC8%40cs.ubc.ca.

[David Poole]

It only lets you add triples and query them efficiently. It uses a Python dictionary, with each subset of {s,v,o} as keys. Given any of the subject, verb or object as given, it return the list of triples that match.

We deliberately don’t add bells and whistles. AIPython is designed as a learning tool, with minimalist implementation that work. The triple store is one of the implementation that is reasonably efficient. 

Python allows concurrent queries, but I wouldn’t trust concurrent additions. (And we don’t support retractions).

David

To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/CALGFikcxTex%2B5ShwngnOVoGDyQ0Pih_ObSLCJDSD4OTSpf5axg%40mail.gmail.com.

[Elisa Kendall]

Hi Michael,

 

I’ve used it - it supports OWL RL, not OWL DL or OWL Full, but includes a datalog rule engine, very useful for all kinds of things. It seems reasonably fast and support has been good. For my purposes, to do some testing with some of our OMG ontologies as well as with parts of FIBO, we needed a bit more semantics than OWL RL, but one of the guys on the team came up with some rules that added cardinality constraints. Lots of functionality has been added since I first got to know them and I’m still learning how their newer features work but if you need OWL support with other capabilities – rules, math, etc. it’s a good choice. I have not done a bake-off with other things though. All of these tools are expanding and adding features constantly, including integration with language models, etc., so it’s hard to keep up

 

For scale they were partnered with AWS Neptune the last time I talked with them, but there may be others now. And they did have capabilities for snapshots to hard storage, and may also have done more with federation since I last investigated.

 

Best,

 

Elisa

--

[Michael DeBellis]

Interesting. So it sounds like you didn't compare it to the usual suspects (Stardog, GraphDB, AllegroGraph)? Any feeling for how limiting the fact that everything is in memory is? I would think that makes scaling more expensive because RAM is more expensive than disk space.  Then again at this point they may both be so cheap it doesn't matter.   Also, I would guess that they aren't as fault tolerant because if they lose power the whole graph is gone and all they have is the last backup. Do you know how accurate that is? The client application is Software as a Service. Any idea how good a fit they are for that? Although if they integrate with Neptune that means they must do SaaS. Any specifics on how they used Neptune to scale? Anyway, that's very useful. Thanks a lot.

Michael

You received this message because you are subscribed to a topic in the Google Groups "ontolog-forum" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/ontolog-forum/U_Yku77l8lo/unsubscribe.
To unsubscribe from this group and all its topics, send an email to ontolog-foru...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/SJ5PPFC295640A5B8F784622B676EEEA591A5C02%40SJ5PPFC295640A5.namprd11.prod.outlook.com.

[Elisa Kendall]

Hi Michael,

 

You asked whether anyone had used RDFox. I have, typically for question answering where certain kinds of analytics, particularly for finance, were needed.  I’ve worked with all 3 of the products you mention, and others such as Top Quadrant’s TopBraid EDG, depending on the use case and what the folks I work with have already purchased. Typically, folks have already made certain purchasing decisions by the time they reach out to me. Or, there are other teams in the same organization doing bake offs, etc. in parallel. But I’m not typically involved unless there are specific constraints / requirements for the use cases that would dictate one product over the other. I haven’t needed to perform the kind of broad, general purpose bake-off you are looking for in other words.

 

RDFox has capabilities my client needed for certain FIBO applications, and where the data was permanently maintained was less important than the ability to perform certain kinds of analyses, for example, complex cash flow and portfolio analyses. There are other features, including how retraction is handled, that RDFox also does well. And up to a point, again depending on the use case, scalability beyond what RDFox can support is less of an issue than whether or not it supports the requirements of particular use cases, often on a subset of the data.

 

I do know that a number of organizations have used RDFox together with Neptune – I think there is a fairly strong relationship there, or at least there was the last time I talked with folks from both organizations. The idea is that the data is maintained in Neptune, but using SPARQL queries to pull relevant subsets into RDFox enables applications like the ones I mentioned. For folks that are already Amazon / AWS shops that are using Neptune, but who have applications that require richer reasoning or analytical capabilities, RDFox is a great option.

 

And I think there is more to the governance capabilities that Oxford Semantics offers than you may have heard, but I would have to go back to Peter Crocker or someone on his team to get an update with more specifics. You might reach out to the Neptune team if you need more detail on the integration, governance, maintenance, and some of the other features you mention.

 

Hope that helps add a little color,

To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/CALGFikf%2B21hm0Yax5xY%3DgGfGgOzXPOgMuA-N%3DHQTKXKDYpYxtw%40mail.gmail.com.

[tkn...@neurocollective.com]

Michael, 

I am working with the in-memory part of Couchbase, very fast, scalable (expand with more hardware), and replicable across data centers. NeuroCollective is a triple store, plus other interesting features. 

Tom Knorr

The NeuroCollective Research

+1 (443) 254-6193

https://www.NeuroCollective.com

--

[Alex Shkotin]

Hi Tom,

" language-agnostic" is a great idea, but usually this means just yet another universal language🤐

What kind of reasoning do we have? This is an important question.

Let's say a reasoning power. RDF reasoning, OWL2 reasoning etc.
What do you think?

Alex

чт, 23 июл. 2026 г. в 11:11, tkn...@neurocollective.com <tkn...@neurocollective.com>:

To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/217010660712381440.0.v2%40titan.email.

[Alex Shkotin]

Hi Michael, no one can replace an expert's opinion. But AI responds instantly.

https://chat.deepseek.com/share/qhluoox2aps9hjmt2q 

I like this "

• Incremental Reasoning: RDFox supports "incremental reasoning." As data is added, changed, or removed, it updates only the logical consequences of that change in real-time, rather than re-processing the entire dataset.

"

The prompt was: "MDB wrote "I have a potential client who is interested in a new RDF triplestore called RDFox. I haven't had time to look at it in detail and would appreciate some feedback if anyone has used or done a serious evaluation of it. All I know now is that it stores all its data in RAM rather than on disk. So I would think that makes it very fast but also brings some challenges regarding scale and fault tolerance." any ideas?" 

Alex

чт, 23 июл. 2026 г. в 01:19, Michael DeBellis <mdebe...@gmail.com>:

I have a potential client who is interested in a new RDF triplestore called RDFox. I haven't had time to look at it in detail and would appreciate some feedback if anyone has used or done a serious evaluation of it. All I know now is that it stores all its data in RAM rather than on disk. So I would think that makes it very fast but also brings some challenges regarding scale and fault tolerance. 

Michael

https://www.michaeldebellis.com/blog

--

[Michael DeBellis]

Elisa, Yes, the potential client that is considering RDFox also mentioned Neptune integration. Thanks for that info, very useful. 

Michael

To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/SJ5PPFC295640A5F3F125C536E9B53E4D97A5C02%40SJ5PPFC295640A5.namprd11.prod.outlook.com.

[Michael DeBellis]

I am working with the in-memory part of Couchbase, very fast, scalable (expand with more hardware), and replicable across data centers. NeuroCollective is a triple store, plus other interesting features. 

Arrgh! Can people stop creating all these new products! Just when I think I understand the landscape there are several ones I've never heard of. Just kidding, thanks for the info. 

Thanks to everyone for sharing their experience. One other thing that is critical IMO is how well they integrate with LLMs. I think they all have toolkits to do that but the only one I've used is the LLM toolkit in AllegroGraph. They work very well. They add triples that point to vectors and the vectors have edges back to domain objects so when you do RAG you generate vectors, usually for properties that have strings for values and those vectors have pointers back to the domain object. So what I do is write SPARQL queries (that's the best way to use the vectors, you have a SPARQL template with parameters and every time you run the LLM you instantiate that query with parameters (the obvious one being the user prompt). Then I add OPTIONAL statements to match common domain object properties. I used to also put the entire query on the copy/paste stack and put the user into Gruff (AllegroGraph's visualization tool) to visualize the results by pasting the generated query. But I found a much cleaner way to do it: now I keep the original query but combine it with results and additional property queries the LLM doesn't need but would be useful for the user to see. I like this better because the old way where I just put the original query on the copy/paste stack (using a wonderful little Python library called Pyperclip) I was calling the LLM twice. For small demos it didn't matter but for a fielded system it was doubling the number of tokens used for each query. If anyone is interested here's the code that does this: https://github.com/mdebellis/SemanticKG-Design/blob/main/data_product/data_product_src/ui_streamforge_chatbot_askmydocuments.py  

Another recent innovation I made was to deal with all the knowledge in the semantic knowledge graph. This was the first thing that was annoying when I started using this: there was a lot of knowledge in the domain in the actual ontology but the LLM couldn't access it because it understands natural language better than Turtle. I also find that a bit ironic and an example of how much things have changed because the assumption in the past was that Turtle was meant for the computer and natural language for the user but LLMs have turned that on its head. I created a nice hack to deal with this problem. It is surprisingly simple but works surprisingly well. I think part of the reason it works well in my current domain (a data catalog prototype) is that the most important info isn't in TBox axioms but ABox property assertions and the latter are easier to just do simple pattern matching. So I match patterns like "?subject :is_downstream_of ?object" and generate strings like "The User Engagement Data Product is downstream of The Clickstream Data Product" Then I generate vectors for that just as I generate vectors for the segments of the documents in the corpus. I'm currently writing a paper on this but if you are interested more detail can be found on a recent blog post: https://www.michaeldebellis.com/post/communicating-ontology-knowledge-with-an-llm-in-graph-rag-the-power-of-simplicity-and-standards  Also, the open source code can be found here: https://github.com/mdebellis/SemanticKG-Design/tree/main/data_product/data_product_src  The two files are generate_embedding_notes_for_properties.py and generate_embedding_notes_for_classes.py

This started out as a question but I find it hard to not talk about what I'm doing!  What I was wondering is if people have experience with other LLM tools in the other Triplestore vendors and can share a bit about how they work as well as strengths and weaknesses? Thanks again for all the replies, very useful.

Michael

https://www.michaeldebellis.com/blog

You received this message because you are subscribed to a topic in the Google Groups "ontolog-forum" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/ontolog-forum/U_Yku77l8lo/unsubscribe.
To unsubscribe from this group and all its topics, send an email to ontolog-foru...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/217010660712381440.0.v2%40titan.email.

[Michael DeBellis]

Thanks Alex, btw, I'm beginning to worry how much longer the first part of your statement "no one can replace an expert's opinion" will be true ;-)

You received this message because you are subscribed to a topic in the Google Groups "ontolog-forum" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/ontolog-forum/U_Yku77l8lo/unsubscribe.
To unsubscribe from this group and all its topics, send an email to ontolog-foru...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/CAFxxRORS61XTU9-344GXJcv0OZa9s8XA67diDsMxrNPWYeuZqA%40mail.gmail.com.

[Alex Shkotin]

There are two ways to verify: by an expert and by an algorithm written by an expert.

GenAI is an algorithm but is not written or verified by any expert.😁

чт, 23 июл. 2026 г. в 18:01, Michael DeBellis <mdebe...@gmail.com>:

To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/CALGFikfununVkPAK1p-bZ%3DbO8Mhkqx-BV4a755uo2ULQik9y4g%40mail.gmail.com.