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, toolCalls, generationId companion stores, plus generate() and abort().

Basic Usage

import { fromLLM, toToolCallRequests } from 'callbag-recharge/ai';
import { effect } from 'callbag-recharge';

const llm = fromLLM({ provider: 'openai', apiKey: 'sk-...', model: 'gpt-4o' });

// Text generation
llm.generate([{ role: 'user', content: 'What is TypeScript?' }]);

// Tool calling
llm.generate(messages, { tools: registry.definitions() });
effect([llm.toolCalls], () => {
    const calls = llm.toolCalls.get();
    if (calls.length > 0) {
      registry.execute(toToolCallRequests(calls));
    }
});

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.
• Generation nonce: generationId is a monotonically increasing Store<number> incremented on each generate() call. Use it to distinguish stale status emissions from previous generations when subscribing to status.
• Tool calling: Pass tools in GenerateOptions to enable function calling. Parsed tool calls accumulate in the toolCalls store. Use toToolCallRequests() to convert to ToolCallRequest[] for toolRegistry.execute().
• 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.