OpenAI 聊天节点

The workhorse chat node

OpenAIChat is the sibling of GeminiChat, but for the chat-completions world - which means it'll happily talk to DeepSeek-R1's reasoning, GPT-4o, Qwen, or whatever model you've pointed the loader at. It returns a plain STRING you can wire into a CLIP text encode, a display node, or downstream text processing. For many users the appeal is the default: cheap DeepSeek-R1 reasoning running through a free-ish ModelScope key is a lot of "let an LLM fix my prompt" power for a few cents.

It ships with the same prompt-optimizer system_prompt as the Gemini side - rewrite input into a high-quality prompt under 200 words - with the same NSFW-tolerant stance baked in. If you don't want the model cheerfully describing explicit content, replace that system prompt with your own. For 2026's LLM-encoder checkpoints, a natural-language prompt optimizer is a genuinely useful stage: these models want instructions, not tag soup.

How it works

The node builds a messages array from your prompt, history, and content parts, then calls chat.completions.create through the loader's OpenAI client. With stream on, tokens arrive incrementally (still collected and returned whole at the end). filter_reasoning is the interesting toggle: reasoning models like R1 emit a separate reasoning_content stream, and with the filter on, only the final answer lands in response_text - the chain-of-thought still prints to the console. Leave it on unless you genuinely want the thinking in your output.

The inputs worth touching

• user_prompt - the main text. Empty if you're sending only content items.
• system_prompt - the "人设" (persona), per the tooltip. Replace the optimizer freely.
• content_part_1/2/3 - three OAI_CONTENT_ITEM slots (text blocks, encoded images, uploaded files). Use a OpenAIContentConnector for more.
• max_tokens - 1024 default, up to 32000. Reasoning models can blow 1024 fast; raise for long output.
• temperature / top_p - sampling; 0.7/0.95 default.
• response_schema_json - structured output via response_format: json_schema with strict: true. Same three accepted formats as the Gemini side: full schema, simplified object, or bare key-value pairs. Note strict mode needs a provider that supports it - not every OpenAI-compatible endpoint does.
• history_json_in / history_json_out - wire the out into the in for multi-turn conversation.
• should_change - false caches the node until an input changes (via IS_CHANGED); true (default) re-runs every queue. Set false for slow, expensive calls.
• retries - 0–5 with backoff.

Outputs

response_text - the model's answer as a STRING. history_json_out - the conversation as JSON.

Troubleshooting

• Timeout on reasoning models. A long R1 chain can exceed the loader's 120s default. Raise the loader's timeout before anything else.
• JSON mode errors - either the model/provider doesn't support strict schema output, or your schema isn't one of the three accepted formats. The node's error message names the parse failure specifically.
• Provider quirks - some endpoints return non-standard shapes; this pack handles the common cases, but exotic hosts can still misbehave.
• Keys and prompts go to whichever endpoint you configured. Keep live keys out of shared workflow files.