Crate sqry_nl 

sqry-nl: Natural Language Translation Layer for sqry

This crate provides translation from natural language queries to sqry commands. It uses a MiniLM-L6-v2-based intent classifier combined with regex-based entity extraction to produce validated, safe sqry commands.

Architecture

The translation pipeline consists of:

1. Preprocessor - Unicode normalization, homoglyph detection, input sanitization
2. Entity Extractor - Regex-based slot filling for symbols, languages, etc.
3. Intent Classifier - MiniLM-L6-v2 ONNX model for intent classification
4. Template Assembler - Maps (intent, entities) to sqry command templates
5. Safety Validator - Whitelist validation, metachar rejection, path guards
6. Translation Cache - LRU cache for repeated queries

Example

use sqry_nl::{Translator, TranslatorConfig, TranslationResponse};

let config = TranslatorConfig::default();
let translator = Translator::load(config)?;

match translator.translate("find authentication functions in rust") {
    TranslationResponse::Execute { command, confidence, .. } => {
        println!("Command: {} (confidence: {:.1}%)", command, confidence * 100.0);
        // Execute the command via sqry CLI
    }
    TranslationResponse::Confirm { command, prompt, .. } => {
        println!("{}", prompt);
        // Ask user for confirmation
    }
    TranslationResponse::Disambiguate { options, prompt } => {
        println!("{}", prompt);
        // Present options to user
    }
    TranslationResponse::Reject { reason, suggestions } => {
        eprintln!("Cannot translate: {}", reason);
        // Show suggestions
    }
}

Safety

All generated commands are validated against a strict whitelist of allowed command templates. The following are always rejected:

• Shell metacharacters (;, |, &, $, etc.)
• Environment variable expansion ($HOME, ${VAR})
• Path traversal (.., absolute paths)
• Write-mode operations (--force, repair, prune)

Feature Flags

• classifier (default) - Enable the MiniLM-L6-v2 classifier. Requires ONNX Runtime. Disable for minimal builds that only need rule-based classification.

Re-exports

pub use cache::CacheConfig;

pub use cache::CacheStats;

pub use error::NlError;

pub use error::NlResult;

pub use translator::Translator;

pub use translator::TranslatorConfig;

pub use types::DisambiguationOption;

pub use types::ExtractedEntities;

pub use types::Intent;

pub use types::OutputFormat;

pub use types::SymbolKind;

pub use types::TranslationResponse;

pub use types::ValidationStatus;

Modules

assembler

Command template assembly.

cache

Translation cache for repeated queries.

classifier

Intent classification using MiniLM-L6-v2 ONNX model.

error

Error types for the sqry-nl crate.

extractor

Entity extraction from natural language input.

preprocess

Input preprocessing for natural language queries.

translator

Main Translator API for natural language to sqry command translation.

types

Core types for the sqry-nl crate.

validator

Safety validation for generated commands.

Constants

VERSION

Crate version for model compatibility checks