Arcane.ai.chat()

This focused page is derived from the mechanically checked full member inventory.

Syntax

Arcane.ai.chat(request)

Parameters

request: AIChatRequest

Return value

Promise<AIChatResult>

Description

Sends a chat request through the configured provider.

Overview

Arcane.ai.chat(request) sends one non-streaming chat request through the configured Ollama or OpenAI provider and normalizes the result. Prefer it when application code should remain provider-neutral. Bind a request to the profile you presented by sending expectedProvider; Arcane then rejects a concurrent provider change instead of silently switching providers.

Parameters

request is a closed AIChatRequest. messages is required and contains 1-128 exact {role, content} records. Roles are system, user, or assistant; each content string is nonempty and at most 131,072 characters, and combined content is at most 512 KiB. Portable optional fields are expectedProvider and format ("json", "", or null). model, options, tools, keep_alive, think, logprobs, and top_logprobs are Ollama-only controls. An omitted local model uses the saved default; OpenAI always uses the saved, account-validated model.

Return value

It resolves to normalized AIChatResult: provider, model, an assistant message, done, doneReason, promptEvalCount, and evalCount. Ollama can also supply bounded message.thinking and message.toolCalls. Response content and thinking are each retained up to 4 MiB.

Availability

The current public method requires ai.inference on desktop Core. It can use the selected local Ollama service or protected OpenAI credential. The current Android projection exposes direct Arcane.ollama.chat(), not this provider-neutral method.

Errors and recovery

Correct invalid request, message, model, format, or context errors before retrying. AI_PROVIDER_CHANGED means refresh Arcane.ai.profile() and let the user confirm the new provider. OpenAI can reject with OPENAI_NOT_CONFIGURED, OPENAI_MODEL_REQUIRED, OPENAI_MODEL_UNAVAILABLE, network, bounded-response, or invalid-response errors. Local model admission errors require a currently runnable admitted model or the managed lifecycle; Arcane never falls back to another provider.

Streaming, cancellation, and events

This API resolves one normalized result and emits no chunk event. The wrapper uses a 130-second renderer timeout and does not expose a per-call signal. Desktop Core can cooperatively cancel the underlying method when it receives a host control frame, including page teardown, but timeout or teardown should still be treated as an unknown completion boundary rather than a retry token.

Example

const arcane = globalThis.Arcane;
const profile = await arcane.ai.profile();
const result = await arcane.ai.chat({
    expectedProvider: profile.provider,
    messages: [
        {role: 'system', content: 'Answer briefly and accurately.'},
        {role: 'user', content: 'Explain the current model in one sentence.'}
    ]
});

console.info(result.message.content);

Reference group

AI and Ollama