Putting a knowledge graph into an LLM is usually not feasible. However, you can put an ontology (structure + taxonomy + rules) that then allow you to query (and update) a knowledge graph that in turn can drive LLM content. In effect, the SHACL becomes your contract between the language model and the knowledge or context graph. This is a long article, but I cover a lot of ground here, including why I believe that such an approach can both dramatically improve accuracy and provide a working environment for dynamic data that can better feed the LLM in its role not as database but as transformer. Thoughts on Queens and cabbages and sailing ships, or at least the role of Steampunk as a programming aesthetic, all of course in … The Ontologist!
Yes, but #SHACL isn’t mandatory. What’s mandatory is understanding what you are trying to achieve —IMHO. Regarding SHACL, as a knowledgeable practitioner, I use it to constrain #SPARQL insert operations associated with special folders mapped to named graphs in a #VirtuosoRDBMS instance. Basically, SHACL is integrated via a folder attribute. The filesystem hook is relevant, and important, due to AI Agents looking to it as the universal interface for both context building and utilization activities.
SHACL provides valuable validation of declared constraints. But it primarily checks conformance to defined shapes and rules. What remains is the question of “behavioral” rules — lifecycle management, authorized transitions, cross-entity constraints — which often go beyond local validation and end up implemented in application logic.
Himanshu Tiwari
1mo
So SHACL or similar standards, such as RDF Schema or OWL, serve as a shared context enabling seamless semantic interactions.
Denis O.
1mo
EXCEPT FOR ONE THING.. any LLM can simply ignore all input context when generating the token trajectory.. so a graph decoupled from internal LLM dynamics or any fancy RAG is no panacea
Charlie Northrup
1mo
Using an ontology plus SHACL as the contract between an LLM and a context graph is a strong way to prevent schema drift and reduce hallucinated structure. One clarification, though, SHACL can enforce conformance (structural validity against declared shapes) but it’s not the same as establishing adequacy for reliance. A node can be perfectly shape-valid and still be stale, unauthorized, misattributed, or contextually unsafe to act upon.
Célian Ringwald
1mo
Pleased to read your article :) I've learnt a lot ! I completly like the "SHACL as a contract" concept ! This is pretty usefull to generate under contraints and validate it afterward. I recently defended a thesis on // SHACL + SML relation extraction : https://hal.science/view/index/docid/5446838 Depending of the context making the use of smaller models and better controlling the cost + sovergnerty of the data is also a subject.
Timothy Cook
1mo
Kurt Cagle, spot on. The LLM cannot reason without a Shape to constrain it. We are finding that SHACL is the perfect 'City Planning' tool for the Graph, but we still need a 'Building Code' for the data before it enters the city. We’re using strict XSD/Archetypes to force that constraint at the Packet level (Ingest). If the packet arrives with H=0 (Zero Entropy), SHACL has much less work to do. Great to see the focus returning to Constraints. https://www.linkedin.com/posts/axius-sdc_the-zero-entropy-data-packet-why-multilevel-activity-7428624538285907969-r0UU
Mikkel Fishman
1mo
I'm going to try https://github.com/Hawksight-AI/semantica And adapt it with shacl to try this out
SHACL as a contract layer between the LLM and the graph is a really clean mental model. I've seen teams burn months trying to get LLMs to write raw SPARQL reliably, when the real move is constraining the interface. The ontology-as-API pattern just makes more sense at scale.
Denis O. Put another way, an empty input string is still an input string. The narrative that comes out will be from latent randomizers due to temperature, in a distribution pattern that is likely some form of noise (probably not fully white noise, depending upon the associated random generator). In this case what comes out is likely oracular: LLM as I Ching, if you will. Of course, this condition is usually trapped early on in a guardrail in most live systems, probably with statements like "Say something profound, wise, or funny." In the event of a supposed zero string prompt.