llm-api-adapter-xai

Official xAI Responses API support for llm-api-adapter.

Installation

Install through the core package extra:

pip install "llm-api-adapter[xai]"

Direct installation is also supported when the core package is managed separately:

pip install llm-api-adapter-xai

Async methods need HTTPX:

pip install "llm-api-adapter[xai,async]"

Synchronous requests remains the default. To opt into the HTTPX synchronous transport, install "llm-api-adapter[xai,httpx]" and pass transport="httpx".

Quick start

import os

from llm_api_adapter.models.messages.chat_message import UserMessage
from llm_api_adapter.universal_adapter import UniversalLLMAPIAdapter

adapter = UniversalLLMAPIAdapter(
    organization="xai",
    model="grok-4.6",
    api_key=os.environ["XAI_API_KEY"],
)

response = adapter.chat(
    messages=[UserMessage("Explain retrieval-augmented generation.")]
)
print(response.content)

Supported models and capabilities

The package deliberately exposes fixed model IDs, not moving aliases: grok-4.5 and grok-4.6.

For grok-4.5 and grok-4.6, xAI cannot disable reasoning: a requested "none" is mapped to the documented minimum and produces a warning.

Structured-output portability

This package requires llm-api-adapter>=0.9.2,<1.0.0 and enforces the same Core portable JSON Schema profile as OpenAI, Anthropic, Google, and Mistral. The profile guarantees that every object is strict, every property is required, optional values are nullable, and only direct, non-recursive local #/$defs/... references are resolved before the request.

xAI's documented immediate schema failures are an additive local overlay, not a replacement for the Core boundary. The adapter rejects boolean property schemas, empty enum or anyOf, minContains/maxContains, tuple items arrays, and unsupported regular expressions before the request. See xAI's structured-output documentation for xAI-specific details.

Use json_schema for parsed JSON only. Use a Pydantic response_model when the final result must also be locally validated and returned as ChatResponse.parsed_model; each nested Pydantic model must use ConfigDict(extra="forbid"). Refusal and incomplete terminal responses set ChatResponse.refusal or ChatResponse.incomplete_reason and leave parsed fields unset. Invalid completed JSON or failed Pydantic validation raises JSONSchemaError. The complete portable vocabulary and examples are in the main Structured Output guide.

Conversations, files, and costs

previous_response is accepted for the shared API, but xAI continuation is intentionally not used: the adapter does not send previous_response_id. Keep and provide the complete messages history for each turn, including assistant tool calls and ToolMessage results.

The package does not expose or send xAI's store option. xAI documents that Responses are stored server-side by default, so configure data retention with xAI when that matters to your application. In particular, the library does not turn on Zero Data Retention (ZDR); enabling ZDR in the xAI Console blocks new Files API uploads and file_id attachments.

PDF URLs are passed to xAI unchanged. For PDF bytes, the adapter uploads a provider-owned file with a 24-hour expiry; it never deletes or changes a URL or file identifier supplied by your application. No OCR or local text extraction is performed.

Attaching a PDF activates xAI's attachment_search tool. That makes the request agentic and adds tool-invocation charges to normal token charges. Storage for an uploaded file is also billed by xAI until it expires. Treat ChatResponse.cost_total as the exact request cost only when xAI returns it; consult xAI billing for storage and any charges not included in that response.

See the official xAI documentation for Responses storage, files and expiry, and file-search pricing.

See the main llm-api-adapter README for the shared API contract and examples.