llm-dialect

llm-dialect = "0.1" # MSRV 1.88

Pure sans-I/O LLM dialect translation: Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses wire formats ↔ a canonical request/response model, plus SSE framer state machines for streamed turns. No async runtime, no HTTP client/server types in the API surface, no wall clock, no randomness — id minting and timestamps are supplied by the caller.

The mapping is the translation core of an LLM gateway (a LiteLLM-proxy-class service): parse whatever dialect your client speaks into one ItemRequest, translate to the ChatRequest your engine consumes, render the engine's canonical chunks back to the caller's dialect.

What it gives an embedder

Put an Anthropic /v1/messages (or OpenAI Responses) surface in front of any chat/completions-class backend without rewriting the translation yourself:

use llm_dialect::{
    canonical::ChatRequest,
    dialect::{anthropic::{req, stream::{chunk_to_sse_events, finalize_stream, StreamState}}, deflate},
    items::ItemRequest,
};

// 1. parse an Anthropic Messages wire body into the canonical model
let items: ItemRequest = req::from_anthropic(&body)?;

// 2. flatten to the internal chat request your engine speaks
let req: ChatRequest = deflate::items_to_chat_request(&items)?;

// 3. per-chunk: canonical chunks back to Anthropic SSE frames
let mut st = StreamState::with_stop_sequences(stops);
for chunk in engine_chunks {
    for (event, data) in chunk_to_sse_events(&chunk, &model, &mut st, &msg_id) {
        write!(out, "event: {event}\ndata: {data}\n\n")?;
    }
}
// one guaranteed terminal pair (message_delta + message_stop)
for (event, data) in finalize_stream(&mut st) {
    write!(out, "event: {event}\ndata: {data}\n\n")?;
}

Enable the axum feature for the shared SSE pump and per-dialect *_stream_response shells (anthropic_stream_response(&stream, …) -> Response), if you're already an axum service and want drop-in handlers.

Crate shape

Guarantees the crate holds

The dialect translation files are pure data transformation. A lint in the integration tests enforces no tokio/reqwest/axum/sqlx/rand/ clock calls outside the sanctioned shell/factory regions, so the boundary above is enforced rather than aspirational.

License

MIT.