Utilities

@obayd/agentic includes a few utility functions, primarily for internal use but potentially helpful for users, especially when implementing the llmCallback.

fetchResponseToStream

This is the most likely utility you might use directly. It's designed to simplify processing streaming responses from LLM APIs that use the Server-Sent Events (SSE) protocol, which is common (e.g., OpenAI, Anthropic in some modes).

Signature:

import { fetchResponseToStream } from '@obayd/agentic';

declare function fetchResponseToStream(
    response: Response // Standard Fetch API Response object
): AsyncGenerator<string>;

Functionality:

• Takes a standard Response object obtained from a fetch call.

• Checks if the response was successful (response.ok). Throws an error with status and body text if not.

• Checks if the response has a body (response.body). Throws if not.

• Reads the response body as a stream (ReadableStream).

• Decodes the stream chunks using TextDecoder.

• Splits the decoded text by lines, handling potential partial lines across chunks.

• Looks for lines starting with data:.

• Attempts to JSON parse the content following data:.

• Specifically looks for text content within the parsed JSON at the path choices[0].delta.content (common in OpenAI-like streams).

• If found, yields the textChunk. If JSON parsing fails or the specific path isn't found, it currently yields the raw data content (after data:) as a fallback.

• Stops when the stream ends or encounters a data: [DONE] message.

Usage Example (in llmCallback):

import { fetchResponseToStream } from '@obayd/agentic';

async function* llmCallback(messages, options) {
    // ... setup fetch request ...
    const response = await fetch(API_ENDPOINT, { /* ... headers, body ... */ });

    // Directly yield from the helper generator
    yield* fetchResponseToStream(response);
}

Note: If your LLM API uses a different streaming format (e.g., newline-delimited JSON, a custom binary format), fetchResponseToStream will likely not work correctly, and you will need to implement your own stream parsing logic within your llmCallback.