Gateway transformations

[remh]

Summary

Add format detection, request transformation infrastructure, and OpenAI-specific preprocessing for gateway integration.

Changes

• Format Detection: Struct-based validation to detect incoming request formats (OpenAI, Anthropic, Google, Bedrock, Mistral) via is_valid_for_format() and validate_or_transform() APIs
• Request Transformation Pipeline: Convert payloads between any supported format through universal message representation with automatic provider defaults (e.g., max_tokens for Anthropic)
• OpenAI Request Transformations: Request preprocessing via OpenAIRequestTransformer builder - reasoning model handling, structured output fallback, provider-specific field sanitization
• Universal Message Transforms: System message extraction and consecutive message flattening for providers that require them
• Media Utilities: URL parsing and base64 conversion utilities for images/PDFs

New Public API

use lingua::{ProviderFormat, TransformResult, TransformError};

// Check if payload is valid for target format (struct-based validation)
if lingua::is_valid_for_format(&payload, ProviderFormat::Anthropic) {
    // Use payload as-is (pass-through)
}

// Full request transformation with provider defaults
let result = lingua::transform_request(&payload, ProviderFormat::Anthropic)?;
match result {
    TransformResult::PassThrough => { /* use original payload */ }
    TransformResult::Transformed { payload, source_format } => {
        // payload has been transformed and defaults applied
    }
}

// Convert to/from universal message format
let messages: Vec<Message> = lingua::to_universal(&payload, ProviderFormat::OpenAI)?;
let anthropic_messages = lingua::from_universal(&messages, ProviderFormat::Anthropic)?;

// Apply provider-specific defaults (e.g., max_tokens for Anthropic)
lingua::apply_provider_defaults(&mut payload, ProviderFormat::Anthropic);#### OpenAI-specific transformations

use lingua::providers::openai::transformations::{OpenAIRequestTransformer, TargetProvider};

// Transform OpenAI requests for downstream providers
OpenAIRequestTransformer::new(&mut request)
    .with_target_provider(TargetProvider::Mistral)
    .with_provider_metadata(metadata)
    .with_supports_streaming(Some(true))
    .transform()?;#### Universal message transforms

use lingua::universal::{extract_system_messages, flatten_consecutive_messages};

// Extract system messages (for Anthropic, Google, Bedrock)
let system_prompts = extract_system_messages(&mut messages);

// Merge consecutive same-role messages
flatten_consecutive_messages(&mut messages);### Supported Formats

Format	Detection	To Universal	From Universal
OpenAI	✓	✓	✓
Anthropic	✓	✓	✓
Google	✓	✓	✓
Bedrock (Converse)	✓	✓	✓
Mistral	✓	✓ (via OpenAI)	✓ (via OpenAI)

Testing

• Payload-based tests in payloads/transformations/ covering provider sanitization, reasoning models, structured output, and cross-format conversion
• Unit tests for all transformation functions

[choochootrain]

looks reasonable at a high level. i had 2 questions that can be addressed outside this PR, just thinking out loud about what end to end behavior we want

[ankrgyl]

just for bookkeeping purposes, undoing my quick approval

[knjiang]

• Gateway transformations #31 : 2 dependent PRs (#38 , #40 ) 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.