# NLI Tool
<p align="center">
<img src="../assets/header.svg" width="960" alt="Physics MCP banner" />
</p>
[Home](../../README.md) · [Architecture](../Architecture.md) · [Configuration](../Configuration.md) · Tools: [CAS](CAS.md) · [Plot](Plot.md) · [NLI](NLI.md)
Natural Language Interface parses free-text physics requests into structured tool calls.
Tool
- `nli_parse`
- Params: `text` (string)
- Returns: `{ intent, args, confidence?, explanation? }`
- Example request:
```json
{"jsonrpc":"2.0","id":"1","method":"nli_parse","params":{
"text":"Plot y = x^2 from -5 to 5"
}}
```
Intents
- CAS: `cas_evaluate`, `cas_diff`, `cas_integrate`, `cas_solve_equation`, `cas_solve_ode`
- Plot: `plot_function_2d`, `plot_parametric_2d`, `plot_field_2d`
- Unknown: `unknown` with an explanation when parsing fails
Operation
- Primary: calls a local LM API (`/chat/completions`) with a physics-aware system prompt when configured.
- Fallback: rule-based parser using regex patterns for common tasks when no LM is configured.
Configuration
- Optional: LM Studio (or any OpenAI-compatible local endpoint). Not required for core calculations.
- Environment (optional): `LM_BASE_URL`, `DEFAULT_MODEL`, and `LM_API_KEY` if your local server requires it.
Why a local LM helps (optional)
- Lower latency and fewer retries on complex requests → faster end-to-end results.
- Uses your GPU (when available) to accelerate parsing.
- Privacy and cost benefits from keeping tokens local.
Quick setup (LM Studio)
1. Install and run LM Studio.
2. Set `LM_BASE_URL` to the OpenAI-compatible endpoint (e.g., `http://localhost:1234/v1`).
3. Set `DEFAULT_MODEL` to your chosen local model.
Schemas & Prompt
- Types and prompt: `packages/tools-nli/src/schema.ts`
- Logic and LM usage: `packages/tools-nli/src/index.ts`
A gentle quip: we keep the parsing Hamiltonian simple-no unnecessary terms, just enough to reach the ground truth.
