In an article for Science in July 2025, Professor Melanie Mitchell of the Santa Fe Institute provided some amusing stories of AI chatbots behaving badly. In one example, a writer asked ChatGPT to critique different essays she’d written and pick one to send to a literary agent. The chatbot responded with an entirely fabricated response.

“When challenged,” Mitchell wrote, “ChatGPT admitted that it could not actually access the essays, and for each one, ‘I didn’t read the piece and I pretended I had.’”

Mitchell pointed out that LLM chatbots are pretrained and encouraged to do role playing. “A useful way to look at these models,” she said, “is as ‘role-players:’ Their vast training on human-generated text has taught them to generate language and behavior in the context of a given role, where the context is set by the user’s prompts.”

Such a telling observation puts chatbots in their place. They are just role playing, which is something businesspeople should always keep in mind. Plain ordinary LLMs aren’t responding to queries deterministically and pulling information from databases to answer questions, and their response accuracy, particularly over a stretch of ground in a business context, is at best low to unpredictable.

While Mitchell’s article was enlightening in its descriptions of the chatbot inaccuracy problem, it was disappointing when it came to identifying a solution. “The solutions to these problems,” she wrote, “are not obvious.” 

What? Those who follow trends in hybrid AI know very well there is a solution. That solution is to get the chatbot to do knowledge graph-based retrieval augmented generation, or RAG. Knowledge graph-based RAG can deliver the accuracy and verifiable answers business users need.

With a good graph RAG solution, the chatbots actually can retrieve trusted, quality information reliably, not only from a database, but from federations of self-describing, contextualized, logically interlinked databases.

Talking to Your Graph: The Basics

On an introductory product information page, Graphwise describes its Talk to Your Graph product as a different kind of chatbot: 

“Talk to Your Graph (TTYG) is a chatbot that allows you to converse with your data and extract factual information using natural language. The chatbot is an example of Graph Retrieval-Augmented Generation (Graph RAG), as it retrieves relevant information from a GraphDB knowledge graph in the form of triples and uses that information to generate informed responses.”

So Graphwise’s TTYG is no ordinary chatbot. Ordinary chatbots are good at interpreting natural language questions and generating natural language answers, period. The quality of the answers, as pointed out in the Science article above, is often open to question. You might get a useful answer, or you might not. Even if the answer seems useful, it’s best to check the veracity, the provenance and the suitability of the answer to the specific question, particularly if you’re relying on the answer for decision making purposes.

Ordinary chatbots, to my mind, are built using monolithic, probabilistic machine learning methods. They’re trained in a statistical way on oodles of unstructured text. 

TTYG takes a hybrid approach: it brings the power of systematic data collection and contextual modeling directly to the end user, using a special, semantic graph form of retrieval generated automation (RAG). 

The front end does use a typical chatbot’s probabilistic approach–OpenAI assistants tap into a deep learning model pretrained to guess what word will come next, but the back end is deterministic, factual and richly, logically connected. That’s the contextualized, richly connected, knowledge graph part.

Regardless of what kind of front end you have, the back end (at the heart of which is GraphDB 11.0) acts as a semantic database management system, including the fact-and-logic handling characteristics of enterprise DBMSes.

Natural Language Knowledge Graph Querying in TTYG

TTYG offers users the ability to select from different search and retrieval methods. Quoting again from the product page: “The chatbot uses several extraction methods — including SPARQL queries, similarity search, simple full-text search, and the ChatGPT Retrieval Connector — to request information from your knowledge graph, based on your query.” Let me explain briefly what these methods entail:

• Similarity search for TTYG finds the closest matches at the word level or document object level.
• Full-text search: Queries retrieve textual data pattern matches from the full set of unstructured text.
• ChatGPT Retrieval GraphDB Connector: Converts RDF to text, which is then vectorized and indexed in a vector database, allowing the relationship-rich RDF to be queried using natural language. 

Key to TTYG is its ability to convert natural language to a semantic graph database query using a language called SPARQL. I first wrote about SPARQL in 2009. It’s a powerful standard query language for knowledge graphs that’s evolved substantially, one that can handle complex queries over very large federations of semantically linked databases.  

But it’s not easy to learn how to write SPARQL queries. So I’m intrigued to try TTYG out and discover more about how this natural language-to-query capability actually works. Such a capability is near and dear to true neurosymbolic or hybrid AI.  I’ll let you know what I find once I have a chance to explore more.