Why Agentic Systems Need Ontologies — Frank Coyle, UC Berkeley

0:01[music]

0:12>> Um, okay, we're going to launch here.

0:15So, my name is Frank Coyle.

0:17Um, I'm I'm an educator and teaching at

0:20Berkeley now. I've been doing this

0:22computer science stuff for

0:24oh, 30, 35 years.

0:26And um

0:28>> [snorts]

0:28>> I'm intro and right now it's kind of a

0:30critical time for uh

0:32poor computer science students. Used to

0:34be the used to be the only game in town.

0:36Degree was a guaranteed job, and now

0:39thanks to AI, it's not. But then again,

0:415,000 people are here. So, AI and and

0:45agents are um seem to be the way to go.

0:48So, the question is how do we leverage

0:50this new universe that we are moving

0:53quickly into. And so, I want to talk

0:55about how agents and ontologies will big

0:58word fit together. But before you do

1:01before I do that, I wanted to um

1:05wanted to give you my uh my educational

1:07philosophy.

1:09And this comes [clears throat] from uh

1:10someone called Sister Corita Kent.

1:13And it was made popular by John Cage,

1:16who is a uh an avant-garde musician.

1:19And

1:21you got to think about this little bit.

1:23Nothing is a mistake.

1:25There is no win.

1:26There's no fail. There's only make.

1:30And more and more today, that's what's

1:32important. Get down and make stuff, and

1:35that's how you're going to learn, not by

1:36necessarily reading.

1:38I'm also a big fan of writing.

1:41My early career was in neuroscience. I'm

1:43kind of coming back into it now that

1:45Agent AI is bringing uh

1:48kind of cognitive science back. But

1:51engage your senses. Get a notebook. Get

1:54a

1:55pen, a pencil. Draw pictures, write

1:58stuff down.

2:00Just don't type because when you type

2:02you when you're typing your brain is

2:04thinking about the letters on the

2:05keyboard. When you're writing in a book,

2:08your whole brain, your your whole all

2:10your all your sensory systems are

2:12engaged and you're going to learn

2:14faster that way.

2:16Okay.

2:18On to our talk.

2:21Agents and ontology. So, there are two

2:23lineages here and I want to talk about

2:25both, give you a little philosophical

2:26background. Um

2:29agents, when did we start talking about

2:31agents? Well, goes goes back to the

2:33early initial days of AI. People like

2:36John McCarthy,

2:38uh

2:39uh uh

2:40uh uh Selfridge, Marvin Minsky, Society

2:43of Mind. People started thinking about

2:45the fact that this new computing

2:47technology was going to lead us into

2:50some kind of artificial intelligence,

2:52which is a term that came

2:54in 1956 when all these characters got

2:56together and tried to figure out where

2:59the future was going. Okay? And the

3:01concept of an agent finally evolved,

3:04things that

3:05perceive and decide and then act and

3:07that's what we're seeing now.

3:09Now, what about ontologies? Well, it

3:10turns out ontologies are not that new.

3:13Okay? It was actually Aristotle who

3:16first came up with the concept of we

3:18need a philosophy of of being. Like,

3:22whoa, kind of heavy. Um but came up with

3:25categories of being and this kind of

3:27relates to what people are doing now

3:29with graph databases and knowledge

3:32representation. And there are a couple

3:34of other people who kind of formalized

3:35it.

3:36Uh

3:37Von Quine was a philosopher and then

3:39this guy Gruber, 1993. And I think this

3:42captures what

3:45knowledge and uh

3:48graph technology really represents. It

3:51is a a formal specification

3:54of a shared conceptualization. And

3:57that's what we want to give to our

3:58agents. We want to give them our concept

4:01our conceptualization of the universe,

4:04our universe, our domains. Okay? And

4:06now, what's happening

4:08is you're getting the convergence of

4:11something that is probabilistic,

4:14the agents, the LLMs, with the

4:18the more formal representations that you

4:20have with ontologies. And so, this term

4:24is now being used you

4:26hearing this a lot, neuro-symbolic

4:28AI. Sounds pretty fancy, but it's really

4:31neural networks tied into

4:35symbolic AI, which rule-based systems

4:39come under that category,

4:41um as do the knowledge graphs that we're

4:44that we're

4:45assembling. And so,

4:48what I'd like to argue is that

4:49neuro-symbolic AI

4:52sort of represents a way to keep the LLM

4:56on its guardrails, because LLMs are by

5:00nature probabilistic.

5:02People worry about hallucinations, but

5:05that's the feature. That's actually a

5:07feature of large language models. It's

5:09who we are. We hallucinate in a way. We

5:13imagine things that may not exist, and

5:15then we turn them into reality.

5:17And that's what large language models do

5:19in

5:19in a way.

5:20Okay? So,

5:23let's just quickly overview what

5:26ontologies are. It's not They're not

5:28complicated. They're basically a

5:29representation of entities and their

5:33relationships to other entities. And

5:35these entities have properties. And this

5:38whole concept of graph databases

5:42arose when

5:45people began to realize that relational

5:48databases

5:49sticking data into tables was too

5:52restrictive. You wanted to add something

5:55new to

5:56a relational database, so you have to

5:58add a new column. Man, I had then then

6:00you have to redo the whole structure.

6:02With a with a graph database, you can

6:04just attach another item. You can just

6:06attach a property. You can attach a

6:07relationship. Okay? So, the question

6:10often arises, okay, I I get it. I need

6:13to have an ontology

6:15to represent in a formal way what my

6:18organization is doing. How do I do it?

6:21Okay? There are a couple of ways you can

6:22approach it. You can have a top-down

6:24approach or a bottom-up approach.

6:26Top-down approach is

6:29you get the experts together and they

6:31sit down and analyze the domain, come up

6:34with the entities. What do we have? We

6:35have purchase orders, we have customers,

6:38we have customer representatives, and

6:41we're going to structure them. They have

6:42properties. These are the relationships.

6:44Okay, that's one way. And this models

6:47what we were doing back in the '80s when

6:49I was involved in expert systems.

6:51Everybody thought expert systems was the

6:53way to do AI. Symbolic AI was the way to

6:56go. Companies rose, millions of dollars

7:00were spent.

7:01Uh the

7:02the Japanese created this uh

7:06future world project in the late '80s.

7:09People in America were my my son was

7:12taking Japanese in school because of

7:14these expert systems. And

7:17but they couldn't scale. They couldn't

7:19scale, and then we went into a kind of

7:21AI winter.

7:23Where did neural networks came come

7:25from? Neural networks were put out there

7:27in the '60s, but they couldn't scale

7:30because we didn't happen to have Nvidia

7:34who was off making GPUs to make make

7:38reality of the of the video games

7:40fantastic, and then someone said, let's

7:43turn these things over to the neural

7:44networks, and of course, that's kind of

7:46why we're here now. So,

7:49the other way you can that that people

7:52are

7:53adding to or creating ontologies is is

7:55from the bottom up. For example,

7:58customer

7:59reactions.

8:01What are the things the customers are

8:02involved in? Wait.

8:04Do you these entities, these

8:05relationships, let's add this to our

8:07ontology. Let's Let's add this

8:09information to the graph. Now, as as a

8:12help,

8:14it's helpful to be aware that there are

8:16existing

8:18taxonomies that people have been working

8:20on for the last 15 to 20 years. Things

8:23like schema.org, which has a whole set

8:25of terms and relationships, so you don't

8:27have to reinvent the wheel. In fact, you

8:30it's to your advantage to use some of

8:33these ontologies. FOAF, Friend of a

8:36Friend, for modeling social networks.

8:40The Dublin Core, which was

8:42an early an early attempt to come up

8:45with terms for describing

8:47uh research papers and books and so

8:49forth. So, there's a whole series of

8:52things. In fact, Wikipedia is based on

8:54an ontology called DBpedia. So, when you

8:57do a search on Wikipedia, it's looking

8:59things up in its giant graph database.

9:02So, this stuff has been out there

9:04underlying a lot of what we already do.

9:07So, take advantage of these things that

9:09already exist.

9:12Okay. Now, what do you do when you build

9:15your ontology? Okay, so what? I know

9:17what these entities are, I know what

9:19their relationships are, they have

9:21properties. How can I do anything with

9:24them? Well, there are other augmenting

9:26technologies, auxiliary technologies.

9:29Things that we call the things like

9:31RDFS, which is a technology, and OWL,

9:33which I'll talk more about. So, these

9:35have

9:36these kind of sit over to the side of

9:39your graph. So, I'm not going to talk

9:41about on I mean ontology is a big word

9:44and it's often confusing and used in

9:45many ways, but think of it as a graph

9:47data structure. Okay?

9:50And you have the entities and

9:51relationships, but you want to apply

9:53some control over them. Or you want to

9:56be able to make inference over them. So,

9:59for example, there is uh

10:02some terms in this technology called

10:04RDFS.

10:05Domain and range. So, if I say

10:09teaches

10:11has a domain of teacher. That means if I

10:14say Bob teaches Scooter in my text, I

10:17can infer that Bob is a teacher.

10:21And if I say all teachers are persons,

10:23then this statement lets me know if I

10:26say Bob teaches Scooter, now I know Bob

10:28is a person, Bob is a teacher. What

10:29about Scooter? If I say teaches has a

10:32range of student, that means the the

10:35right side of the verb, then Scooter is

10:37a student. And now I have this extra

10:39information into my system.

10:43OWL also has a series of

10:47of of properties that allow you to make

10:50some inferences. So, a transitive

10:52property transitive property says, if

10:55Sue is an ancestor like ancestor is a

10:58transitive property. If Sue is an

11:00ancestor of Mary and Mary is an ancestor

11:02of Ann, then Sue is an ancestor of Ann.

11:07Okay? This was not initially into my

11:10graph system, but with applying these

11:13functional properties, I can then

11:16add and augment the system with this

11:18extra data. So, that's very useful.

11:21Then there's some

11:23properties called functional properties,

11:26which means only one.

11:28So, has father

11:31is a functional property. You can only

11:33have one father. You can only have one

11:35mother. That is a functional property.

11:38Okay? So, that's that can serve as a

11:41constraint. So, when

11:44if you say Bob is my Bob is Jim's

11:48father, BB is Jim's father,

11:51well,

11:53the inference here is that Bob and BB

11:55are two ways of representing the same

11:57individual because

11:59that is a functional property. Can only

12:01have one. So, these derivations and

12:04constraints that don't sit in the graph,

12:06they sit sort of on the side

12:09and they can help

12:11as we're going to see, I'm going to

12:12propose,

12:14when we deal with agents, how they can

12:16they can help us out. So,

12:18what about agents?

12:21Everybody's talking about agents now and

12:23everybody's talking about loops. Loops,

12:26loops, loops everywhere.

12:28Loops have been around for a long time.

12:31Back in the '60s, people were debating

12:34who has the best programming language?

12:35Fortran or COBOL? No, mine is better.

12:37No, mine is better. Oh, you don't know

12:39anything. You don't know what you're

12:40talking about.

12:42Bohm and Jacopini in 1966 came out and

12:45said, "Okay, there is no real difference

12:47in programming languages if they have

12:49three aspects.

12:51Sequence. I can put statement A,

12:53statement B, statement C. Fine. I have

12:56conditionals. I can have if then.

12:58And the last piece, I have a loop.

13:01If I have a loop, if I have iteration,

13:03if I take these three things,

13:05the the language is what's called Turing

13:08complete. Can do any can compute

13:10anything that a

13:13can be computed by computational devices

13:17from the work of Alan Turing.

13:19Okay?

13:20And now we're seeing this in agentic AI.

13:24Agents are now have loops. Loops give us

13:26the last piece in the equation of giving

13:29us

13:31a technology that is capable of

13:35doing anything that computational

13:37devices can do.

13:40The danger though of loops is that they

13:43can break.

13:45If you're If you're a programmer, you

13:46know, you've all go into infinite loop.

13:49Not good.

13:51Loops can drift as agents start talking

13:55to each other, things get all go off off

13:58the rails. And

14:00loops can cost you money.

14:02Token counts crank up as the loops

14:05continue. So, you don't you need to be

14:08careful, okay? But in a way we are

14:11revisiting some of the early stuff with

14:13symbolic AI. I would argue we're going

14:16back to the world of expert systems.

14:20Which is the symbolic part of the whole

14:21thing. So, I want to show you a little

14:23example using Claude

14:26agent. So, little code here. Don't get

14:29scared, but

14:31I know nobody does Python anymore, but

14:33you got to look at what the agent's

14:35giving you and you got to you got to

14:37move in and and manipulate it. So,

14:39here's a here's a loop while true,

14:41classic Python loop. Okay? And so, we

14:44have a client. So, we're actually So,

14:46the first little chunk here that you

14:48see, r e s p, the response, this is just

14:51some code where we have a model and we

14:55have uh

14:57we have a prompt, that's part of part of

14:59the messages,

15:00and we have a tool,

15:03and we're we're asking the LLM to

15:08solve this problem using a tool. Now,

15:11here's the here's the catch.

15:14LLMs can't do anything. All they can do

15:17is give us the next word with a high

15:19probability. Amazingly, we can now have

15:22these conversations it, but they can't

15:23do anything. But, we can give it a tool,

15:27and we can give it

15:29what we want, and say,

15:32"How do you think this tool can help us

15:34get what we want?" And then the LLM will

15:37set up the parameters,

15:41and come back to us, and say, "Okay,

15:44here's my response.

15:46I can't execute this tool, but I know

15:48what the input parameters are. I know

15:50what your context is. I know what your

15:53prompt is. So, here is the call that you

15:56need to make

15:58of the tool, because I can't do it. I'm

16:00the LLM. I'm just locked in this box.

16:03Okay? So, the second box the second

16:05chunk is

16:08stop reason. So, stop reason means

16:12the LLM has stopped for some reason.

16:15The The reason here is that it can't do

16:17anything, and if the reason is tool use,

16:20ah, now it's time. Let's go execute that

16:22tool. So, that second line, get tool. It

16:25takes the response, which is

16:28formulating the the parameters, and

16:30triggering the action.

16:32Okay. Now, I have this stuff in red

16:34here. This is where I think

16:36the LLMs and other uh

16:40I'm sorry, not LLMs.

16:41The ontologies and stuff can come in.

16:43So,

16:45if you look down there,

16:48after the the tool is called, it said

16:50tool runs. This is where

16:54ontologies could come in.

16:56The tool's going to give us information.

16:59We put the information in a form

17:02that our

17:04our our our validator can use, and think

17:07about the validator as operating with

17:10this these ontologies about our domain,

17:13then we can

17:16make some sense

17:17of whether the response of the LLM is

17:22reasonable. So, this is the loop. Call a

17:25tool, check the stop reason.

17:28If it's a reasonable result, then let's

17:31go with it. If it's not reasonable, go

17:33back to the LLM. Say, "Oh, this is this

17:35is not working." Or get a human in the

17:38loop. But the idea is

17:42to surround the input with checks. Now,

17:44I've got this

17:45something that you that you should be at

17:48least taking a look at if you're doing

17:49some of this coding is something called

17:51Pydantic. Pydantic is a way to specify

17:56the types of what you want the types of

17:58the parameters to be. Those of you who

18:01who do know Python, know Python is a

18:04unstructured type language. So, you can

18:06have a variable x = 20, x = hello, no

18:10problem. There's no typing.

18:12Pydantic adds typing to that. So, you

18:14want to

18:15check your types with Pydantic and then

18:18check your results with the ontology.

18:23So, Pydantic at the door, ontology at

18:25the ledger, and pure agents and by the

18:28way, your agents should try to have no

18:30side effects.

18:32That helps the whole logic. Meaning,

18:34they're not running off doing something

18:36that they're they're changing they're

18:38changing things in the database not yet.

18:40You want to run them through the

18:41ontology first and make sure that works.

18:44Okay.

18:45I only got an I've got I've got another

18:46I've just a short time. I'm going to try

18:48to show you some of the things that um

18:51that you can

18:53some logical constructs from from

18:55something called OWL, the uh

18:58the web object language for for objects.

19:01So, you have these functional

19:03properties, disjoint properties. I'll

19:05just put these you can look at the

19:06slides, but essentially the errors it

19:09can catch. Look over in the the

19:11right-hand column. A second refund on

19:14the same order

19:15is a is

19:17is a problem. But ontologies could catch

19:19it, whereas it's it's very tricky to do

19:21that in in English. A payout sent to the

19:24support desk instead of the buyer. Okay?

19:27You can catch that with an owl disjoint

19:29property where customer and support rep

19:31are two separate entities. Okay?

19:34Uh one of may a made-up value like

19:38probably shipped. You can specify

19:42you must have certain kinds of value.

19:44So, uh the status

19:47paid, shipped, or refunded, nothing

19:49else. And when you're in the pure text

19:51world, this can get this can get funky

19:54because the the LLMs are again

19:57probabilistic and

19:58um

20:00return some crazy stuff. Okay. Uh so,

20:03really what the point I want to make

20:05here is use these re- you can have a

20:08reasoner built on ontology to check keep

20:11the LLM on track, have guardrails to

20:16keep it honest. Okay?

20:18And for the guardrails, I'm referring to

20:21these concepts re- these support

20:24technologies with RDFS and owl.

20:28And

20:30my my bottom line is and nothing is a

20:32mistake, there's no win, no fail, only a

20:35make.

20:37Okay. Feel free to reach out to me

20:39coil@burkly.

20:40I've got a I've got a

20:41I've got a little website

20:42codesupreme.ai. I'm a big fan of if

20:45you're John Coltrane has a has a some

20:48jazz called uh

20:49called Love Supreme. So, I've named my

20:52site

20:53Code Supreme. And if you go there, I've

20:54got some music and it's all good. Okay.

20:57Thanks very much. 20 minutes.

20:59>> [applause]

21:15[music]

21:17>> Woo!