# Legacy LLM interface

Source: https://openflow.run/en/docs/nodes/llm/
Language: en
Product version: 0.1.0-beta.77

Reference for existing LLM definitions and their message, input, and result contracts. This is not an Add Node option.

The Add Node panel does not offer a standalone LLM node. This page documents existing LLM definitions and their CLI/MCP interface. For a new model task, choose AI Agent in the panel or a model action actually supplied by your app catalog.

Use the LLM interface for an explicit set of messages.

## Messages and model

Configure the model and template defaults. Templates interpolate named inputs such as `{{input}}`. History supplied through `messages` precedes template messages and is not interpolated.

```json
{
  "op": "node.add", "as": "summary", "type": "llm", "name": "Summarize a ticket",
  "config": {
    "model": { "model": "deepseek-v4-flash" },
    "template": [{ "role": "user", "content": "Summarize in one sentence: {{input}}" }]
  },
  "inputs": { "input": { "kind": "value", "value": "The customer cannot complete checkout." } }
}
```

The model must be available through the deployment’s configured model service. Bind only the data required for the request.

An LLM Task interpolates the Start input into a message template and returns one text output.

**Input**
{"text":"The customer cannot complete checkout."}

**Expected result**
A generated one-sentence summary; exact wording depends on the model.

**Before running**
Select a model available to the deployment and configure its model service before running.

[Download workflow](https://openflow.run/docs/examples/summarize-text.openflow.json)

## Defaults and overrides

Configuration values for `model`, `template`, and `messages` are saved defaults. Input bindings with those names take precedence. Editing a default preserves an existing binding.

Use the `default` source to restore inheritance. Use `unset` to suppress a default explicitly. These differ from a literal JSON null value.

## Results

The public authoring interface manages the result port. Results default to text; `config.resultSchema` requests a structured result contract. The service validates outputs against that contract before committing the Task result.

A schema constrains shape, not factual correctness. Review generated content for the task’s requirements. Exact wording is model-dependent and is not established by the offline preview.

## Failures

LLM supports ordinary [Task retries and handled errors](https://openflow.run/en/docs/concepts/errors-and-retries/index.md). Retries repeat the model Task with the same inputs and may produce different generated content or incur additional model usage.

## Sources

*   [docs/authoring/node-authoring.md](https://github.com/oomol-lab/open-flow/blob/e25a738d170322734572a0b8a1323cb00b9d355a/docs/authoring/node-authoring.md)
*   [docs/control/contracts/control-api.md](https://github.com/oomol-lab/open-flow/blob/e25a738d170322734572a0b8a1323cb00b9d355a/docs/control/contracts/control-api.md#node-configuration-semantics)
*   [packages/open-flow/src/workbench/browser/runtime/editor/addNodeOptions.ts](https://github.com/oomol-lab/open-flow/blob/e25a738d170322734572a0b8a1323cb00b9d355a/packages/open-flow/src/workbench/browser/runtime/editor/addNodeOptions.ts)