soil-guard-cli

Soil Guard CLI

A focused CLI tool designed to demonstrate the implementation of LLM Guardrails. This project acts as a "Soil Health Service" assistant, using parallel input and output validation to ensure AI responses remain on-topic, safe, and formatted correctly.

Overview

The Soil Guard CLI simulates a professional agricultural assistant. It uses a middleware-inspired approach to guardrails:

• Input Guardrail: Intercepts the user's prompt to ensure it is strictly related to soil science, plants, or agriculture.
• Output Guardrail: Scrutinizes the LLM's response to prevent "hallucinated" chemical recommendations or non-professional advice.
• Parallel Execution: Runs the guardrail check and the primary LLM completion simultaneously to minimize latency.

Key Features

• Topic Enforcement: Rejects queries about unrelated topics (e.g., weather, stocks, general tech) using a lightweight classifier model.
• Safety Filtering: Detects and flags potentially harmful or unauthorized agricultural recommendations.
• TUI Experience: Uses commander.js for command handling and chalk for color-coded status reporting in the terminal.
• Type-Safe Validation: Uses Zod to define and enforce schemas for guardrail results and structured LLM outputs.

Suggested Capabilities

Guarded Querying (Input Guardrail)

The core feature will be a "Topic Jail." When a user submits a query via the CLI, the program first evaluates if the query is relevant to soil health.

• If relevant: It passes the query to the LLM.
• If irrelevant: It uses chalk.yellow or chalk.red to inform the user that the request is out of scope (e.g., "I only discuss soil, not the weather").

Soil Data Analysis (Structured Output)

Using the logic from your other projects, you can feed the tool specific soil metrics (pH levels, Nitrogen, Phosphorus, Potassium).

• Validation: The guardrail can check if the LLM's output contains a specific "Confidence Score" or "Safety Warning" before displaying it.
• Formatting: It can use chalk to color-code soil health indicators (e.g., pH 5.0 shows up in red for high acidity).

"Human-in-the-Loop" Verification

Following the pattern of your other AI projects, the CLI can prompt the user for confirmation before "finalizing" a recommendation.

CLI Flow:

• User inputs: soil-health analyze --ph 6.5
• Guardrail verifies query.
• LLM generates a recommendation.
• Output Guardrail checks for "hallucinated chemicals."
• TUI displays the advice and asks: Is this advice safe to follow? (y/n)