Email replies with Gmail approval and a knowledge base

Email replies with Gmail approval and a knowledge base

An n8n workflow that answers incoming email from your own documents, with a person approving every reply. One part, which you start by
hand, reads a Google Drive folder and stores the documents in a Qdrant collection. The other part watches a mailbox over IMAP and
summarizes each new email with DeepSeek. An AI agent searches the collection and writes a reply, and Gmail sends the draft to an approver.
The approver types an answer, and a classifier reads it as approved or declined. An approved reply goes to the sender by SMTP. A declined
one is rewritten from the feedback and sent for approval again.

What is inside

• workflow.json: the workflow as n8n exports it, 31 nodes. 22 of them run and nine are sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• Create collection sent a POST with an empty filter, which does not create a collection. It now sends a PUT with a vector size of 1536
  and the Cosine distance. An error in that step does not stop the run.
• QDRANTURL and COLLECTION are now <YOUR_QDRANT_URL> and <YOUR_COLLECTION>. The author's test folder in Get folder and the
  approver's address in Gmail are now <YOUR_DRIVE_FOLDER_ID> and <YOUR_APPROVER_EMAIL_ADDRESS>.
• The shared OpenAI node named gpt-4o-mini, which OpenAI lists for shutdown on 14 January 2027. It now names gpt-5.6-terra. Two agents
  with a tool use this node, and gpt-6-luna, the listed replacement, calls tools on Chat Completions only with reasoning effort none.
• DeepSeek Chat Model had no model set, so it took the node's default, deepseek-chat. DeepSeek announced that name as discontinued on
  24 July 2026. The node now names deepseek-flash.
• In the two prompts of Email Summarization Chain, the email text has its curly braces doubled. The prompt is a template, and one stray
  brace in an email would make the step fail.
• Markdown and the approval message use the plain-text part when an email has no HTML part. Before, such an email gave an empty summary.
• The sticky notes name the new placeholders. The note about the approval step said the approver needs a Gmail address. n8n's source only
  asks for one valid address, so the note now says what the step does.
• Every credential link is gone. Select your own.
• Nothing else. The prompts, the categories, the splitter and the flow are as they were.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Qdrant instance, on Qdrant Cloud or your own server, with its address and an API key.
• An OpenAI API key, for the embeddings, the two agents and the classifier, and a DeepSeek API key for the summary.
• A Google Drive folder of Google Docs that answer the questions you get.
• An IMAP mailbox that receives the questions and SMTP details to send the replies. A test mailbox is a good start.
• A Gmail account that sends the approval emails, and the address of the approver. It can be your own.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. In Create collection and Refresh collection, replace <YOUR_QDRANT_URL> with your Qdrant host, with its port if it needs one,
   and <YOUR_COLLECTION> with a collection name of your choice. Add a Header Auth credential with the name api-key and your Qdrant
   key as the value, and select it in both nodes.
3. Add a Qdrant API credential and select it in Qdrant Vector Store and Qdrant Vector Store1. Replace <YOUR_COLLECTION> in both
   with the same name.
4. Add an OpenAI credential and select it in OpenAI, Embeddings OpenAI and Embeddings OpenAI1. Add a DeepSeek credential and
   select it in DeepSeek Chat Model.
5. Add a Google Drive OAuth2 credential and select it in Get folder and Download Files. In Get folder, replace
   <YOUR_DRIVE_FOLDER_ID> with your folder.
6. Run the first part once from When clicking ‘Test workflow’. It creates the collection, clears it, then reads the folder, splits
   each document and stores the pieces.
7. Add an IMAP credential and select it in Email Trigger (IMAP). Add an SMTP credential and select it in Send Email. Add a Gmail
   OAuth2 credential, select it in Gmail and replace <YOUR_APPROVER_EMAIL_ADDRESS> there.
8. Send a test question from another address, one that your documents answer. Open the approval email, follow its link and type Ok.
   Check the reply that arrives, then save and press Publish at the top of the canvas (or Shift+P).

How a run goes

The first part, started by hand:

1. When clicking ‘Test workflow’ starts Create collection and Refresh collection. The second sends a delete-points request with an
   empty filter before each fill.
2. Get folder lists the files in your folder, and Download Files downloads each Google Doc as plain text.
3. Default Data Loader and Token Splitter cut each file into pieces of 300 tokens with an overlap of 30.
4. Embeddings OpenAI1 turns the pieces into vectors, and Qdrant Vector Store1 stores them.

The second part, for each new email:

1. Email Trigger (IMAP) picks up a new unread email from the INBOX and marks it read.
2. Markdown turns its HTML into Markdown in a field named data.
3. Email Summarization Chain asks DeepSeek Chat Model for a summary of at most 100 words.
4. Write email is an agent on OpenAI. It searches the collection through Qdrant Vector Store and writes a reply of under 100
   words, the body only.
5. Edit Fields keeps the reply in a field named email.
6. Gmail sends the approver an email with the original message, the draft and a link. The run waits there for a typed answer.
7. Text Classifier reads the answer as Approved or Declined.
8. For Approved, Send Email sends the draft by SMTP to the sender of the email, from the address that received it, with the
   subject Re: and the original subject.
9. For Declined, Email Reviewer rewrites the draft as HTML from the draft and the feedback. Edit Fields keeps the new version,
   and Gmail asks again. The loop goes on until someone approves.

Change the look

• Which documents it knows: the folder in Get folder, and the pieces in Token Splitter.
• How a reply reads: the system message of Write email. How a rewrite looks: the system message of Email Reviewer.
• What counts as approval: the categories in Text Classifier.
• The length of the summary: the two prompts in Email Summarization Chain.
• The models: OpenAI for both agents and the classifier, and DeepSeek Chat Model for the summary. Change the embedding model in
  both embedding nodes and the vector size in Create collection together.
• The approval email: Gmail.

Good to know

• Not run: Mindket made no call to n8n, IMAP, SMTP, Gmail, OpenAI, DeepSeek, Qdrant or Google Drive. It ran the changed expressions on
  its own machine and read n8n's source for how the nodes behave.
• OpenAI serves three nodes, both agents and the classifier. The model page for gpt-5.6-terra lists function calling, and the
  page for gpt-6-luna says Chat Completions supports it only with reasoning effort none.
• deepseek-flash thinks before it answers by default, so a summary can take a little longer.
• Email Summarization Chain is set to read a binary file, and there is none. The email text reaches the model through the prompt.
• One answer decides. Anything the classifier reads as approved sends the reply to the original sender, so a vague answer can send a
  draft. Test with your own address first.
• Gmail has no time limit set, so n8n waits with no end for each email that gets no answer. Each one keeps a run open.
• Send Email uses the to header of the incoming email as its from address. An email sent to several people, or one with a display
  name, may fail or go out from the wrong address.
• Both embedding nodes use text-embedding-3-small, which has 1536 dimensions, the size Create collection now asks for.
• Refresh collection clears the collection before each fill. Qdrant's reference page does not say what an empty filter matches, so
  try it on a throwaway collection first.
• The agent's tool is named company_knowladge_base. The spelling is the original's and does no harm.
• Your data: the email text goes to DeepSeek. The summary, the searches and the approver's feedback go to OpenAI. The documents go to
  OpenAI to be embedded and are stored in Qdrant. The approval email carries the original message through Gmail.
• A message can carry text meant to steer the agent. A person approves every reply, so read each draft before you answer.
• Credentials: the JSON holds none.