Build a shopping chatbot in four minutes with GraphDB Talk to Your Graph 2.0

A fast path to a “shopping chatbot” is now practical with GraphDB’s Talk to Your Graph 2.0: load product data into a GraphDB 10.8 repository, then let a graph-augmented LLM answer plain-English questions by translating them into SPARQL (SPARQL is referenced as “Sparkle” in the transcript) over the graph. The key payoff is speed—setting up the repository, importing RDF/Trig data (including a named-graph ontology), creating an agent, and running real queries can be done in under four minutes.

The workflow starts with an RDF file in TriG syntax containing product descriptions, categories, and review/rating information on a 1–5 scale. The file also includes a small ontology stored in a named graph: it defines two classes and several properties, plus product categories modeled as subclasses of a broader Product class. That ontology placement matters because Talk to Your Graph uses it to understand what classes and properties exist when generating queries.

On the GraphDB side, a new repository named “products” is created with full-text search enabled. The RDF/Trig import is configured so that triples land in the correct named graphs: the ontology goes into the named graph specified in the source file, while the rest of the product data goes into the default graph. A quick SPARQL check confirms the data loaded correctly—for example, a query for products whose descriptions contain “mirror” returns two matching items (a makeup product and a sync product).

Next comes the agent setup in the GraphDB “Talk to Your Graph 2.0” interface. An OpenAI API key must be configured (the transcript notes it can be set in a GraphDB config file). The agent is created for the “products” repository with full-text search available and a cap of 10 results. Crucially, the agent is given the ontology’s named-graph identifier so it can generate accurate SPARQL queries aligned with the schema.

Once created, the shopping bot answers natural-language requests about the graph. For watch-related questions, response times are reported as roughly 8–13 seconds (with some wait time trimmed from the demonstration). It can list watches and prices, restructure results on demand (e.g., separate men’s and women’s watches), and work with review data—such as identifying a product with the most reviews (a Rolex) and producing a summary of those reviews instead of dumping the full text.

The system also supports transparency and learning: an “explain” option returns the reasoning behind the SPARQL query it ran and shows the generated query itself, helping users understand how plain-English prompts map to graph queries.

Finally, the bot handles more decision-oriented shopping questions. It finds the least expensive watch whose average review score exceeds 3, and it can compute sale pricing—turning an $89.99 price into a $53.99 cost under a 40% discount scenario. The overall message is that with minimal configuration—data import, ontology in a named graph, and an agent wired to an LLM—customers can query graph-backed product catalogs in plain English without learning query syntax.