> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/sdgm/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/sdgm/_mcp/server.

# How Is Kumo Relational Different from an LLM or Text-to-SQL?

> Understand when to use relational prediction, language generation, or database retrieval

Kumo Relational predicts unknown or future outcomes from structured relational data. A large language model (LLM) generates or interprets language, while text-to-SQL translates a natural-language question into a database query that retrieves existing records.

These capabilities complement one another rather than solving the same problem.

## Compare the workloads

| User question                                              | Best fit               | What it does                                                                    |
| ---------------------------------------------------------- | ---------------------- | ------------------------------------------------------------------------------- |
| "How many orders did this customer place last month?"      | **SQL or text-to-SQL** | Retrieves and aggregates facts already stored in the database.                  |
| "Summarize this customer's recent support conversations."  | **LLM**                | Interprets text and generates a natural-language response.                      |
| "Will this customer place an order in the next 30 days?"   | **Kumo Relational**    | Predicts an unknown outcome from historical, temporal, and cross-table context. |
| "Which products is this customer most likely to buy next?" | **Kumo Relational**    | Ranks related entities using relational behavior and context.                   |

## Why the relational graph matters

Flattening several tables into text or serializing them as JSON does not preserve database structure efficiently. Kumo Relational uses table metadata, primary-foreign key relationships, timestamps, and sampled neighborhoods to reason over the relational graph.

Its prediction objective is also different from next-token generation. Kumo Relational is designed for classification, regression, forecasting, and link prediction tasks expressed in PQL. See [Prediction Types](/rfm/prediction-types).

## Use Kumo Relational with an LLM

An application can use each component for what it does best:

1. An LLM interprets the user's request or manages the conversation.
2. Application logic formulates or selects an approved PQL query.
3. Kumo Relational returns the prediction.
4. The LLM explains or presents the result in natural language.

Keep prediction execution and language generation distinct so the application can validate the PQL task and preserve the returned values. For architectural background, see [Kumo Relational: A Relational Foundation Model](/research/kumorfm-paper).