@qualve/llm

Core LLM task framework for Qualve. Provides the LLMTask base class that all LLM provider adapters extend.

If you want all providers out of the box, install @qualve/llms instead.

Setup

Requires Node.js v23+.

npm install @qualve/llm

Usage

Import alongside a provider to register the llm task type:

import "@qualve/anthropic"; // or @qualve/openai, @qualve/googleai

Writing a custom provider

import { LLMTask } from "@qualve/llm";

export default class MyProvider extends LLMTask {
	static id = "my-provider";
	static name = "My Provider";
	static models = ["my-model-v1"];
	static capabilities = {};

	// Required: implement these abstract methods
	async uploadFile (filepath, { mimeType, contents }) { /* ... */ }
	async getFile (filepath) { /* ... */ }
	async deleteFile (filepath) { /* ... */ }
	async listFiles () { /* ... */ }
	async createStream () { /* ... */ }
}

LLMTask.register(MyProvider);

API

LLMTask

Extends the base Qualve Task class with LLM-specific functionality:

• Provider dispatch — LLMTask.create() routes to the registered provider based on task.llm
• File management — Upload, retrieve, and manage files on the provider
• Streaming — handleStream() writes streamed responses to disk with backpressure handling
• Prompt helpers — file.describe() generates prompt text describing task I/O files
• Thinking levels — Normalized across providers via thinkingLevels and per-provider levelMap
• Stop reasons — Normalized stop reasons (COMPLETE, MAX_TOKENS, ABORTED, UNKNOWN)

Abstract methods (providers must implement)

| Method | Description | | --- | --- | | uploadFile(filepath, { mimeType, contents }) | Upload data to the provider | | getFile(filepath) | Retrieve a previously uploaded file, or null | | deleteFile(filepath) | Delete a previously uploaded file | | listFiles() | List all uploaded files | | createStream() | Create the streaming API call; returns { stream, transformChunk, onChunk?, onFinish? } |

Optional overrides

| Method | Description | | --- | --- | | getStatus(chunk) | Extract a human-readable status from a streaming chunk | | countTokens() | Count input tokens for a dry run |