fromLLM()

Creates a unified reactive source for LLM inference via any OpenAI-compatible endpoint.

Signature

function fromLLM(opts: LLMOptions): LLMStore

Parameters

Parameter	Type	Description
opts	LLMOptions	Provider configuration (provider, baseURL, apiKey, model).

Returns

LLMStore — Store<string> with status, error, tokens companion stores, plus generate() and abort().

Basic Usage

import { fromLLM } from 'callbag-recharge/adapters';
import { effect } from 'callbag-recharge';

const llm = fromLLM({ provider: 'ollama', model: 'llama4' });

effect([llm], () => {
    console.log(llm.get()); // accumulating response...
  });

llm.generate([{ role: 'user', content: 'What is TypeScript?' }]);
// llm.status.get() → "active"
// llm.get() → "TypeScript is..."

Options / Behavior Details

• Provider-agnostic: Works with OpenAI, Ollama, Anthropic (via proxy), Vercel AI SDK, or any OpenAI-compatible endpoint.
• No hard deps: Uses fetch + SSE line parsing. No SDK imports required.
• Auto-cancel: Calling generate() while streaming aborts the previous generation.
• Token tracking: tokens store populated on stream completion (when usage data is available).
• Status: Uses WithStatusStatus enum (pending → active → completed/errored) for consistent lifecycle tracking.
• Persistent source: This is a long-lived store backed by state(). It does not send callbag END — lifecycle is managed imperatively via generate()/abort(), not via stream completion. Do not wrap with withStatus() or retry() — use the built-in .status and .error companions instead.