llm_core 0.3.2

Core abstractions for LLM (Large Language Model) interactions. Provides common interfaces, models, and utilities used by LLM backend implementations.

example/README.md

llm_core examples

llm_core has no backend of its own — it defines the abstractions every backend implements. Install it alongside a backend package (llm_ollama, llm_vllm, llm_chatgpt, llm_claude, llm_gemini, llm_llamacpp) and program against LLMChatRepository.

See provider_agnostic_example.dart for a runnable example.

Programming against the interface

import 'package:llm_core/llm_core.dart';

Future<String> summarize(LLMChatRepository repo, String model, String text) async {
  final response = await repo.chatResponse(model, messages: [
    LLMMessage(role: LLMRole.user, content: 'Summarize:\n\n$text'),
  ]);
  // `content` is nullable: a refusal or filtered response arrives as a
  // successful response with no content.
  return response.content ?? '';
}

Any backend can be passed in, so swapping providers is a one-line change at the composition root.

Shared options

LLMChatOptions carries generation, reasoning, tool, structured-output, timeout, retry, cache and metrics settings. Each backend maps them onto its own wire format, and drops or translates what its API does not accept:

const options = LLMChatOptions(
  temperature: 0.2,
  maxOutputTokens: 512,
  think: true,
  responseFormat: JsonFormat(),
);

Finish reasons

Always check finishReason before treating a response as valid output — LLMFinishReason.refusal in particular arrives as a successful response with empty or partial content:

switch (response.finishReason) {
  case LLMFinishReason.refusal:
    throw StateError('Provider safety classifiers declined the request.');
  case LLMFinishReason.length:
    throw StateError('Truncated — raise maxOutputTokens.');
  case LLMFinishReason.contentFilter:
    throw StateError('Output was filtered.');
  default:
    print(response.content ?? '');
}