Classify messages in n8n with Chat Trigger and Basic LLM Chain

1. Prepare an instance and a new workflow

You need an n8n instance with editor access. This exercise starts at workflow creation; Cloud setup, self-hosted installation and maintenance are outside its scope. The official manual quick-start route is Personal → Create workflow. Name the workflow “Message triage exercise”.[1]

• Use one text message at a time and synthetic inputs.
• Add Chat Trigger to the canvas and connect its main flow to Basic LLM Chain.
• Keep Make Chat Publicly Available off in Chat Trigger and use the manual chat interface. Keep Load Previous Session set to Off for this exercise.

[1][2]

2. Connect a model and configure credentials

Add a DeepSeek Chat Model sub-node at the model connection of Basic LLM Chain. Following n8n’s DeepSeek credentials guide, sign in to your own provider account, create a key on its API keys page and place it in the n8n credential’s API Key field. No key is included here. Keep secrets out of prompts, public exports and screenshots.[3][4][5]

Select a model listed for your current account under Model rather than copying someone else’s identifier. This node loads available models dynamically. If none appear, check credentials and account access first. Your balance, eligibility, request costs and successful connectivity have not been verified.[4][5]

3. Define input and classification rules

• Set Basic LLM Chain’s Prompt to Take from previous node automatically. Its parameter documentation requires chatInput from the preceding node.
• Add a System message under Chat Messages with the original rule below.
• Return a single category name in this exercise and keep Require Specific Output Format off. Enabling it later requires an output parser connection.

[3]

Original System message: “Classify only the user message. Login, password or session problems are Login issue. Questions about product capabilities are Feature question. A message without a clear issue is Insufficient information. Return exactly one of those three category names, with no explanation and no facts the user did not provide.” This is an exercise rule, not a guarantee of model compliance.[3][6]

4. Test one message at a time and record actual output

Send each input below through the editor’s manual chat interface. Check whether Chat Trigger outputs chatInput and whether the chain receives the same text. Then inspect the chain output and chat reply. Testing executes the workflow and uses the configured model, so first check your execution allowance and provider billing terms.[2][3]

Article sources[6]

Record each input, model, prompt version and actual output. Compare the target label and any extra text. Rerun the same inputs after changing the model or prompt. These four cases are an introductory check and cannot establish accuracy, reliability or production readiness.[6]

5. Trace the data flow before public access

• For No prompt specified, first check that Prompt is valid and nonempty. Automatic input mode also needs the upstream chatInput field.
• If the chain produces a reply but chat shows something unexpected, inspect Response Mode. A simple flow can use When Last Node Finishes to return the last executed node’s data; Using Response Nodes needs configured response nodes.
• Basic replies use text or output. Renaming these fields in a later step can make Chat Trigger send the whole object.
• Keep chat private until actual test results, authentication and execution usage have been checked. The public chat Authentication setting supports access restrictions.

[3][2]

Verification scope

Checked: the manual workflow creation route, Chat Trigger access and response settings, Basic LLM Chain prompt and input parameters, DeepSeek Chat Model and credential configuration, and official guidance on testing multiple inputs. Each source has a check date and saved content hash.[1][2][3][4][5][6]

Not executed: account signup, Cloud provisioning, self-hosted installation, key creation, model requests, workflow runs, public chat or cost measurement. Inputs and target labels are original exercise designs. Output quality, Chinese-language performance, free eligibility and production reliability remain unverified. Follow the interface and models available to your current instance and account.[2][4][6]