The Open Standard for Generative UI
import { identityMessageFormat, type MessageFormat } from "../types/messageFormat";
import type { StreamProtocolAdapter } from "../types/stream";
import type { ChatLLM } from "./types";
export interface FetchLLMOptions {
/** Endpoint that accepts POST'd messages and returns a streaming Response. */
url: string;
/** Stream protocol adapter for parsing the response body (e.g., agUIAdapter, openAIAdapter). */
streamAdapter: StreamProtocolAdapter;
/** Wire-format conversion for outgoing messages. Defaults to identity (canonical Message). */
messageFormat?: MessageFormat;
/** Extra headers merged into the request. */
headers?: Record<string, string>;
/** Override fetch implementation (for tests, custom auth wrappers, etc.). */
fetch?: typeof fetch;
/** Extra fields merged into the request body (e.g. `model`) */
body?: Record<string, unknown>;
}
/**
* Generic HTTP-based LLM adapter. POSTs an AG-UI `RunAgentInput`-shaped body
* (`{ threadId, runId, messages, tools, context }`, messages in the chosen wire
* format) to `url` and returns the streaming `Response` for downstream processing.
*/
export function fetchLLM({
url,
streamAdapter,
messageFormat = identityMessageFormat,
headers,
fetch: customFetch,
body,
}: FetchLLMOptions): ChatLLM {
const fetchImpl = customFetch ?? globalThis.fetch.bind(globalThis);
return {
send: ({ threadId, messages, signal }) => {
const runId = crypto.randomUUID();
return fetchImpl(url, {
method: "POST",
headers: {
"Content-Type": "application/json",
...headers,
},
body: JSON.stringify({
...body,
tools: [],
context: [],
threadId,
runId,
messages: messageFormat.toApi(messages),
}),
signal,
});
},
streamProtocol: streamAdapter,
};
}