Company knowledge email replies with approve or disapprove

Company knowledge email replies with approve or disapprove

An n8n workflow that answers incoming email from your own company documents. An AI agent looks facts up in a Qdrant
collection and drafts a reply. It then emails the draft to you with an Approve and a Disapprove button. Approve sends the
reply to the sender. Disapprove makes the agent write another draft.

What is inside

• workflow.json: the workflow as n8n exports it, 12 nodes and 5 sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The summary step ran deepseek/deepseek-r1:free on an OpenAI Chat Model node pointed at OpenRouter. OpenRouter's
  public model list, read on 12 November 2026, no longer shows that model. The step now uses OpenRouter Chat Model, the
  sub-node n8n has for OpenRouter, set to deepseek/deepseek-v4-flash. That model is listed, at 0.09 dollars per million
  input tokens and 0.18 per million output tokens on that day.
• The agent's tool was named company_knowladge_base. It is now company_knowledge_base. The agent's model node, once
  called OpenAI, is now OpenAI Chat Model.
• The Qdrant collection read =COLLECTION, and Send Draft read YOUR GMAIL ADDRESS. They are now
  <YOUR_QDRANT_COLLECTION> and <YOUR_GMAIL_ADDRESS>.
• Two sticky notes were rewritten in a few lines each. One told the reader to tokenize documents first, and the other
  claimed that only Gmail can send and wait for an answer. n8n's Send Email node can do it too.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Qdrant instance with a collection that holds your company documents as vectors. This workflow only reads the
  collection. Fill it first with another workflow, using OpenAI's text-embedding-3-small, the model that
  Embeddings OpenAI uses when it has no model set.
• One mailbox you can read over IMAP and send from over SMTP.
• A Gmail account for the approval emails.
• An OpenRouter API key and an OpenAI API key.

Set it up

1. In n8n, create a workflow, open its menu, choose Import from File and pick workflow.json.
2. Add an IMAP credential to Email Trigger (IMAP), the same account's SMTP credential to Send Email, and a Gmail
   credential to Send Draft.
3. Add an OpenRouter credential to OpenRouter Chat Model. Add an OpenAI credential to OpenAI Chat Model and
   Embeddings OpenAI. Add a Qdrant credential to Qdrant Vector Store.
4. In Qdrant Vector Store, replace <YOUR_QDRANT_COLLECTION> with the name of your collection.
5. In Send Draft, replace <YOUR_GMAIL_ADDRESS> with the Gmail address that should receive drafts.
6. OpenAI retires gpt-4o-mini on 14 January 2027. Pick gpt-5.6-terra in OpenAI Chat Model, because the agent calls
   a tool and OpenAI says gpt-6-luna supports function calling in Chat Completions only with reasoning effort set
   to none.
7. Save the workflow, switch it on and send a test email to the mailbox.

How a run goes

1. Email Trigger (IMAP) watches the inbox. It fires for each unread email and marks it read, which are its own
   defaults. The message body sits in textHtml.
2. Markdown turns that into Markdown and puts it in data.
3. Email Summarization Chain asks OpenRouter Chat Model for a summary of at most 100 words. The chain is set to read a
   binary file, and there is none, so the email text reaches the model through the prompt, which holds {{ $json.data }}.
4. Set Email copies the summary into a field called email.
5. Write email is an agent on OpenAI Chat Model. Its instructions ask for a short, professional business reply, with
   the body only, in HTML with a few simple tags. It gets the summary. Its one tool, Qdrant Vector Store, searches your
   collection through Embeddings OpenAI and hands back matching passages.
6. Send Draft emails the original message and the draft to your Gmail address through two buttons, one to approve and
   one to disapprove, then waits.
7. Approve? checks data.approved. On true, Send Email sends the latest draft as HTML to the sender, under the
   subject Re: plus the original subject. On false, Set Email and Write email run again for a new draft, and
   Send Draft asks again.

Change the look

• Tone and length of the reply: the instructions in Write email.
• What the agent may look up: the description of the tool in Qdrant Vector Store.
• How long the summary is: the two prompts in Email Summarization Chain.
• The models: OpenRouter Chat Model and OpenAI Chat Model.
• The button names: the options of Send Draft.
• How long it waits: add Limit Wait Time in the options of Send Draft.

Good to know

• The facts below come from pages and n8n source read on 12 November 2026.
• This workflow does not load your documents into Qdrant. With an empty or mismatched collection, the agent finds
  nothing and writes from the summary alone.
• The collection must hold vectors made by the same model as Embeddings OpenAI, which uses text-embedding-3-small
  here. A collection built with another model returns poor matches.
• A reply is written from the summary, so a detail the summary drops cannot reach the reply. Read each draft against the
  original.
• Send Draft has no time limit. The run waits until you press a button. If you never do, nothing is sent, and the
  waiting run stays in n8n until you delete it.
• Each Disapprove writes a fresh draft from the same summary, so nothing stops you from asking again and again.
• The buttons are links to your n8n. You must be able to open that address from where you read the approval email.
• The email text sits inside a prompt template, so a message with curly braces may make the summary step fail.
• The trigger takes unread mail only. An email you open in your own mail app first is marked read and is skipped.
• Models: OpenAI's deprecations page gives 14 January 2027 as the shutdown date of gpt-4o-mini, with gpt-6-luna as
  the replacement. The agent here uses a tool, which is why the setup steps point to gpt-5.6-terra.
• Billing: OpenRouter and OpenAI bill each run, and your mail provider may cap how many messages you send.
• Your data: the email text goes to OpenRouter and its model provider, the summary and the matched passages go to OpenAI,
  and the approval email goes through Gmail.
• Credentials: the JSON holds none.