# DSPy

Run predictions using self-hosted DSPy programs

[DSPy](https://github.com/stanford-oval/dspy) is an open-source framework for programming—rather than prompting—language models. DSPy enables you to build interpretable and modular LLM-powered agents using Python functions, structured modules, and declarative signatures, making it easy to compose, debug, and reliably deploy language model applications.

With DSPy in SteelEngine, you can:

* **Run custom predictions**: Connect your self-hosted DSPy server and invoke prediction endpoints for a variety of natural language tasks.
* **Chain of Thought and ReAct reasoning**: Leverage advanced DSPy modules for step-by-step reasoning, multi-turn dialogs, and action-observation loops.
* **Integrate with your workflows**: Automate LLM predictions and reasoning as part of any SteelEngine automation or agent routine.
* **Provide custom endpoints and context**: Flexibly call your own DSPy-powered APIs with custom authentication, endpoints, input fields, and context.

These features let your SteelEngine agents access modular, interpretable LLM-based programs for tasks like question answering, document analysis, decision support, and more—where you remain in control of the model, data, and logic.

## Usage Instructions

Integrate with your self-hosted DSPy programs for LLM-powered predictions. Supports Predict, Chain of Thought, and ReAct agents. DSPy is the framework for programming—not prompting—language models.

## Tools

### `dspy_predict`

Run a prediction using a self-hosted DSPy program endpoint

#### Input

| Parameter          | Type   | Required | Description                                                                                          |
| ------------------ | ------ | -------- | ---------------------------------------------------------------------------------------------------- |
| `baseUrl`          | string | Yes      | Base URL of the DSPy server (e.g., [https://your-dspy-server.com\\](https://your-dspy-server.com\\)) |
| `apiKey`           | string | No       | API key for authentication (if required by your server)                                              |
| `endpoint`         | string | No       | API endpoint path (defaults to /predict)                                                             |
| `input`            | string | Yes      | The input text to send to the DSPy program                                                           |
| `inputField`       | string | No       | Name of the input field expected by the DSPy program (defaults to "text")                            |
| `context`          | string | No       | Additional context to provide to the DSPy program                                                    |
| `additionalInputs` | json   | No       | Additional key-value pairs to include in the request body                                            |

#### Output

| Parameter   | Type   | Description                                                     |
| ----------- | ------ | --------------------------------------------------------------- |
| `answer`    | string | The main output/answer from the DSPy program                    |
| `reasoning` | string | The reasoning or rationale behind the answer (if available)     |
| `status`    | string | Response status from the DSPy server (success or error)         |
| `rawOutput` | json   | The complete raw output from the DSPy program (result.toDict()) |

### `dspy_chain_of_thought`

Run a Chain of Thought prediction using a self-hosted DSPy ChainOfThought program endpoint

#### Input

| Parameter  | Type   | Required | Description                                                                                          |
| ---------- | ------ | -------- | ---------------------------------------------------------------------------------------------------- |
| `baseUrl`  | string | Yes      | Base URL of the DSPy server (e.g., [https://your-dspy-server.com\\](https://your-dspy-server.com\\)) |
| `apiKey`   | string | No       | API key for authentication (if required by your server)                                              |
| `endpoint` | string | No       | API endpoint path (defaults to /predict)                                                             |
| `question` | string | Yes      | The question to answer using chain of thought reasoning                                              |
| `context`  | string | No       | Additional context to provide for answering the question                                             |

#### Output

| Parameter   | Type   | Description                                                     |
| ----------- | ------ | --------------------------------------------------------------- |
| `answer`    | string | The answer generated through chain of thought reasoning         |
| `reasoning` | string | The step-by-step reasoning that led to the answer               |
| `status`    | string | Response status from the DSPy server (success or error)         |
| `rawOutput` | json   | The complete raw output from the DSPy program (result.toDict()) |

### `dspy_react`

Run a ReAct agent using a self-hosted DSPy ReAct program endpoint for multi-step reasoning and action

#### Input

| Parameter       | Type   | Required | Description                                                                                          |
| --------------- | ------ | -------- | ---------------------------------------------------------------------------------------------------- |
| `baseUrl`       | string | Yes      | Base URL of the DSPy server (e.g., [https://your-dspy-server.com\\](https://your-dspy-server.com\\)) |
| `apiKey`        | string | No       | API key for authentication (if required by your server)                                              |
| `endpoint`      | string | No       | API endpoint path (defaults to /predict)                                                             |
| `task`          | string | Yes      | The task or question for the ReAct agent to work on                                                  |
| `context`       | string | No       | Additional context to provide for the task                                                           |
| `maxIterations` | number | No       | Maximum number of reasoning iterations (defaults to server setting)                                  |

#### Output

| Parameter       | Type   | Description                                                        |
| --------------- | ------ | ------------------------------------------------------------------ |
| `answer`        | string | The final answer or result from the ReAct agent                    |
| `reasoning`     | string | The overall reasoning summary from the agent                       |
| `trajectory`    | array  | The step-by-step trajectory of thoughts, actions, and observations |
| ↳ `thought`     | string | The reasoning thought at this step                                 |
| ↳ `toolName`    | string | The name of the tool/action called                                 |
| ↳ `toolArgs`    | json   | Arguments passed to the tool                                       |
| ↳ `observation` | string | The observation/result from the tool execution                     |
| `status`        | string | Response status from the DSPy server (success or error)            |
| `rawOutput`     | json   | The complete raw output from the DSPy program (result.toDict())    |
