n8n-nodes-qwen-extractor

An n8n community node that extracts structured data from conversation transcripts and free text using Qwen with strict JSON Schema output.

Features

• Any schema — paste any JSON Schema (orders, leads, invoices, tickets…). The model is forced to match it via response_format: json_schema with strict: true.
• Live model dropdown — loaded from the credential's /models endpoint, with JSON Schema-capable models listed first and the rest clearly flagged. Falls back to a built-in list when the endpoint has no such route.
• Fails before it spends tokens — empty input, missing model, and malformed schemas are rejected before the API call.
• No silent empty results — an empty response, a content-filter block, or a reply that landed in reasoning_content raises a clear error instead of returning {}.
• Token and duration tracking — a summary line in the output pane, including cached tokens.
• Clean output — only the fields in your schema, flat at the root.

Requirements

• Node.js >= 22.16 (required by @n8n/node-cli — see .nvmrc).
• An API key from Alibaba Cloud Model Studio (DashScope), or any OpenAI-compatible gateway serving Qwen.

Installation

Follow the community node installation guide and enter the package name n8n-nodes-qwen-structured-extractor.

Credentials

This node uses the Qwen Structured Extractor API credential owned by this package (type qwenStructuredExtractorApi). It is not shared with the qwenApi credential from n8n-nodes-azbot-qwen-model — both packages can be installed on the same n8n instance without interfering.

| Field | Description | | --- | --- | | API Key | Your sk-... key from Alibaba Cloud Model Studio (DashScope) or another OpenAI-compatible gateway | | Base URL | Defaults to https://dashscope-intl.aliyuncs.com/compatible-mode/v1 |

The credential's Test button calls GET /models.

Choosing a model — IMPORTANT

This node always sends response_format: {"type": "json_schema", "strict": true}. Per the Alibaba documentation, JSON Schema mode is supported only by:

• the qwen3.7-plus series
• the qwen3.7-max series
• the qwen3.8-max series

Every other model (qwen-turbo, qwen-plus, qwen-max, and the older aliases) supports json_object at best and will answer this node with a 400. The dropdown still lists them, marked "may reject"; embedding, image, and speech models are hidden entirely because they cannot serve a chat/completions call.

Usage

Drag the Qwen Structured Data Extractor node onto the canvas, attach the credential, then fill in the parameters below.

Parameters

• Model Name or ID — loaded from /models. Switch the field to Expression mode to use an ID that isn't listed.
• System Prompt — what you want the model to do.
• User Input / Conversation History — an expression pointing at the conversation, e.g. {{ $json.chat_log }}.
• JSON Schema Output — the schema the output must conform to.

Options

| Option | Default | Effect | | --- | --- | --- | | Auto Ensure JSON Keyword | true | Appends the word "JSON" to the system prompt. Required by json_object mode; json_schema mode does not need it. | | Disable Thinking Mode | true | Sends enable_thinking: false. DashScope rejects non-streaming calls to Qwen3 reasoning models without it, and thinking mode makes JSON output less reliable. Turn off only if a model rejects the parameter. | | Include Execution Trace | false | Adds a meta key with per-step durations and the token breakdown. If your schema already has a meta field, the trace moves to _meta instead of overwriting it. | | Include Raw Request/Response | false | Adds the outgoing payload and raw response to meta. Only applies when Include Execution Trace is on. | | Output Key Name | (empty) | Wraps the extracted data in a single key instead of emitting it flat. Required if your schema's root is an array. | | Schema Name | extracted_data | Identifier for the schema inside the API payload. | | Temperature | 0 | Keep at 0 for consistent extraction. |

Output

By default, exactly the fields in your schema:

{ "amount": 2, "size": "L", "phone": "0900..." }

Token counts are deliberately kept out of the output so the item shape stays valid under additionalProperties: false. Read them from the output pane summary instead:

2/2 items in 2.4s | ~1087 tokens (prompt 912 / completion 175) | 640 cached

Turn on Include Execution Trace to get meta.tokens for cost accounting.

Qwen-specific tuning notes

• Don't put per-item expressions in the System Prompt. DashScope's implicit context cache matches on a shared prefix (≥1024 tokens, ~2000 for Qwen3.7) and bills cache hits at 10%. The system prompt must be identical across items for the cache to hit — everything that varies belongs in User Input.
• The node never sends max_tokens. Alibaba's docs state it truncates the JSON mid-output when structured output is on. This is deliberate — don't add it.
• Strict mode: Alibaba recommends additionalProperties: false and listing fields in required. The node warns when they're absent but still runs.
• Thinking models can leak reasoning into content. The node strips prose around the JSON automatically, but leaving Disable Thinking Mode on is the reliable fix.

Troubleshooting

| Symptom | Cause | | --- | --- | | 400 mentioning enable_thinking | The model requires this parameter on non-streaming calls. Turn on Disable Thinking Mode. | | 400 mentioning response_format / json_schema | The model does not support strict JSON Schema. Switch to qwen3.7-plus, qwen3.7-max, or qwen3.8-max. | | 400 messages must contain the word json | The system prompt lacks the word "json". Turn on Auto Ensure JSON Keyword. | | Qwen returned no content | Empty response — a content filter, or the model replied in reasoning_content. The error message names the cause. | | Qwen returned an array at the top level | Your schema's root is an array. Set Output Key Name to wrap it. | | "User Input / Conversation History" is empty | The expression doesn't point at the field holding the conversation. | | No model selected | Open the dropdown and pick a model — the node ships no default. | | JSON parse failure with finish_reason=length | The output was cut off. Shorten the input or simplify the schema. | | Credential doesn't appear | Restart n8n after updating the package. |

Development

nvm use                # Node >= 22.16, see .nvmrc
npm install
npm run dev            # n8n at http://localhost:5678, hot reload
npm run build
npm test
npm run lint

Node < 22.12 makes n8n-node build/lint/dev fail with ERR_REQUIRE_ESM — that's why engines and .nvmrc exist.

License

MIT