Auto-Analyze, Classify, and Summarize Gmail with n8n, OpenAI RAG, and Google Sheets

Gmail new-mail trigger → extract & clean → RAG vector retrieval classification → Google Sheets storage → self-learning taxonomy update → Telegram audio summary notification

Who

Sales, support, or ops teams who need to automate high-volume Gmail processing with a custom taxonomy

You get

New emails are automatically extracted, classified via RAG against historical examples, summarized, and logged to Google Sheets; an audio summary is pushed to Telegram; new categories auto-extend the taxonomy with admin alerts

You need

n8n · gmail · google-sheets · telegram

Steps

1. Configure trigger and credentials

   Create the following credentials in n8n: Gmail OAuth2 (for trigger and message retrieval), Google Sheets OAuth2 (for storage and taxonomy dataset), OpenAI API (for chat, embeddings, and audio generation), Telegram API (for notifications). Add a Gmail Trigger node set to poll every minute. Add a Fetch full Gmail message node to retrieve the full message using messageId.

2. Extract and normalize email fields

   Add a Normalize email fields (Edit Fields/Set) node to extract and map: id, date, sender (from.value[0].name), email (from.value[0].address), subject, content (using html.removeTags() to strip HTML). Ensure plain-text output for downstream AI processing.

3. Prepare the RAG taxonomy knowledge base (manual first run)

   Create two sheets in Google Sheets: Sheet 1 'Email Log' with columns date, email, name, subject, summarized text, category, subcategory, keywords; Sheet 2 'Tagging Samples' with columns id, subject, email_text, category, subcategory, keywords. Add a Get row(s) in sheet node to read the full Tagging Samples table. Add a Prepare taxonomy documents node to concatenate subject and email_text as documents. Add a Create embeddings (OpenAI) node to generate vector embeddings. Add a Vector store — taxonomy index node in insert mode with memoryKey=email_tagging. Add a Vector store retriever (RAG tool) node in retrieve-as-tool mode (topK=1) for the Agent to call. Manually click 'Execute workflow' once to populate the vector store before activation.

4. Configure the AI analysis Agent (RAG classification and summarization)

   Add an Analyze email with RAG agent (AI Agent) node, connect OpenAI Chat Model (gpt-4.1-mini) as the language model, and connect Vector store retriever as a tool. In the Agent's text field, input the email variables. In options.systemMessage, paste the system prompt. The prompt instructs the model to: use retrieved vector-store examples as the sole source of truth for taxonomy; prefer existing categories and only create new ones when no reasonable match exists; output strict JSON appending summary, short_message, category, subcategory, keywords, confidence, new_category, new_subcategory; and perform internal validation before responding.

   Execution prompt · Source text。AI Agent node text input field (User Message); preserve n8n expression syntax

   Copy prompt

   =sender: {{ $json.sender }}
   email: {{ $json.email }}
   date: {{ $json.date }}
   subject: {{ $json.subject }}
   content: {{ $json.content }}

   Execution prompt · Source text。AI Agent node options.systemMessage field (System Prompt)

   Copy prompt

   =You are an email analysis, classification, and summarization assistant operating inside an automated workflow.
   
   You will receive ONE incoming email object as input.
   
   Your task is to analyze the email and RETURN THE SAME OBJECT with additional structured fields appended.
   
   --------------------------------------------------
   PRIMARY OBJECTIVES
   --------------------------------------------------
   
   1. Summarize the email content in ONE short paragraph (maximum 3 sentences).
   
   2. Classify the email using semantic context retrieved from "Simple Vector Store1".
   
   3. The retrieved examples represent the OFFICIAL taxonomy and are the PRIMARY SOURCE OF TRUTH for categories and subcategories.
   
   4. Extract up to 5 meaningful keywords representing the main topics of the email.
   
   5. Provide a confidence score between 0 and 1 indicating classification certainty.
   
   --------------------------------------------------
   CLASSIFICATION RULES (STRICT)
   --------------------------------------------------
   
   - ALWAYS prefer categories and subcategories found in retrieved examples.
   - DO NOT invent new categories or subcategories unless no reasonable semantic match exists.
   - Retrieved examples have higher authority than general model knowledge.
   - If similarity is weak, still choose the closest existing category and LOWER the confidence score.
   - Only create a new category or subcategory when NO retrieved example reasonably matches the email topic.
   
   Decision priority order:
   1) Retrieved vector store examples (highest priority)
   2) Semantic meaning of the email
   3) General reasoning (last resort)
   
   --------------------------------------------------
   TAXONOMY CONSISTENCY CHECK (MANDATORY)
   --------------------------------------------------
   
   After selecting a category and subcategory:
   
   1. Compare the selected category against ALL categories present in the retrieved examples from "Simple Vector Store1".
   
   2. If the selected category DOES NOT appear in retrieved examples from knowledge_base and you generated new one:
      set new_category = true
      otherwise set new_category = false.
   
   3. If the category exists but the selected subcategory DOES NOT appear in retrieved examples from knowledge_base and you generated new one:
      set new_subcategory = true
      otherwise set new_subcategory = false.
   
   4. This decision MUST be based ONLY on retrieved examples, not on general knowledge.
   
   --------------------------------------------------
   CONFIDENCE SCORING
   --------------------------------------------------
   
   0.90–1.00 → strong semantic match with retrieved examples
   0.70–0.89 → clear topic with minor ambiguity
   0.50–0.69 → weak similarity but reasonable classification
   0.00–0.49 → uncertain classification
   
   Lower confidence when retrieved similarity is weak or ambiguous.
   
   --------------------------------------------------
   OUTPUT REQUIREMENTS (CRITICAL)
   --------------------------------------------------
   
   - Output MUST be valid JSON ONLY.
   - DO NOT include markdown, explanations, comments, or extra text.
   - DO NOT remove, rename, or modify any existing input fields.
   - Preserve all original fields exactly as received.
   - Append ONLY the following fields:
   
   {
     "summary": "short paragraph summary",
     "short_message": "1 long line summary",
     "category": "main category",
     "subcategory": "specific subcategory",
     "keywords": ["keyword1","keyword2","keyword3"],
     "confidence": 0.0,
     "new_category": false,
     "new_subcategory": false
   }
   
   --------------------------------------------------
   NEW CATEGORY FLAGS
   --------------------------------------------------
   
   - new_category = true ONLY if the selected category is NOT present in retrieved examples.
   - new_subcategory = true ONLY if the category exists but the subcategory is NOT present in retrieved examples.
   
   --------------------------------------------------
   FINAL VALIDATION STEP (MANDATORY)
   --------------------------------------------------
   
   Before responding, internally verify:
   
   ✓ Output is valid JSON
   ✓ No missing commas or trailing text
   ✓ All original input fields remain unchanged
   ✓ Exactly the required fields were appended
   ✓ No additional fields were added
   ✓ Boolean fields contain true or false values only
   ✓ confidence is a number between 0 and 1
   
   If validation fails, correct the output before responding.

5. Parse results and store in Google Sheets

   Add a Parse AI analysis result (Edit Fields/Set) node to parse the Agent's JSON output with parseJson(), mapping to: sender, email, date, subject, content (summary), short message, category, subcategory, keyword (keywords.join()), confidence, new category, new subcategory. Then add a Store processed email (Google Sheets) node in append mode to write into the Email Log sheet, with field mappings matching the Sheet 1 column structure from Step 3.

6. Self-learning taxonomy update and admin notification

   Add a Check for new taxonomy items (If) node to test whether new category or new subcategory equals "true". On the true branch: add Load taxonomy dataset to read existing Tagging Samples; add Get next taxonomy ID (Code node, Python) to compute max id; add Append new taxonomy entry to write the new category/subcategory into Tagging Samples; add Notify admin about taxonomy update (Telegram) to send a text alert.

   Execution prompt · Source text。Telegram node Notify admin about taxonomy update text field

   Copy prompt

   =New category/subcategory added:
   {{ $json.category }} / {{ $json.subcategory }}

7. Generate audio summary and push to Telegram

   On the false branch of Check for new taxonomy items (or in parallel), add a Generate audio summary (OpenAI) node with resource set to audio and input from the Store processed email summarized text. Then add a Send audio summary to Telegram node with operation set to sendAudio, binaryData=true, caption containing sender, formatted date, and subject, and fileName set to "email bite". Replace chatId with your actual Telegram user or group ID.

You are setting up an n8n workflow to automatically analyze, classify, and summarize Gmail emails. Please confirm and execute in this order:
1. Verify you have: an n8n instance, a Google account (Gmail + Sheets), an OpenAI API key, and a Telegram Bot token with chat ID.
2. Create four credentials in n8n: Gmail OAuth2, Google Sheets OAuth2, OpenAI API, Telegram API.
3. Create the Email Log and Tagging Samples sheets in Google Sheets per the instructions, and seed Tagging Samples with initial classification examples.
4. Import or rebuild the node graph from workflow.json, paying special attention: copy the AI Agent systemMessage and text fields verbatim; use memoryKey=email_tagging consistently across vector store nodes; execute manually once to populate the vector store before activation.
5. Replace [YOUR_CHAT_ID] placeholders in Telegram nodes with real IDs.
6. Activate the workflow and send a test email. Verify: the email appears in Email Log, categories are assigned correctly, and Telegram receives the audio summary.