Daily urgent email alert with a local Ollama model

Daily urgent email alert with a local Ollama model

An n8n workflow that reads the last day of your inbox over IMAP with a Python script and mutes the usual promotions and job alerts. It
asks a model running in Ollama on your own machine which threads need action. When there are any, it sends a short report with a
table of the urgent threads to Telegram, and to an ntfy topic if Telegram fails. On a quiet day it sends nothing.

What is inside

• workflow.json: the workflow as n8n exports it, fourteen nodes that all run and no sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The made-up values YOUR_EMAIL@example.com, YOUR_APP_PASSWORD, YOUR_IMAP_HOST, YOUR_NTFY_TOPIC, YOUR_TELEGRAM_CHAT_ID and
  YOUR_OLLAMA_MODEL are now placeholders in angle brackets, such as <YOUR_APP_PASSWORD>.
• The example mail in the prompt came from a real service address. It is now Example Accounts <no-reply@example.com>.
• The credential link on the Telegram node was removed. Select your own.
• Nothing else. The script, the pattern lists, the prompts, the model settings and the layout of the message are as they were.

Before you start

• A self-hosted n8n with the Execute Command node turned on. n8n Cloud does not have it, and n8n 2.0 switches it off by default, so
  remove it from NODES_EXCLUDE.
• Python 3 and a POSIX shell on the machine that runs n8n. In Docker, the command runs inside the n8n container.
• An IMAP mailbox with its host, user and an app password. For Gmail, Google's help says an app password needs 2-Step Verification.
• Ollama on the same machine, listening on http://127.0.0.1:11434, with a chat model pulled.
• A Telegram bot token and your chat id.
• A hard-to-guess ntfy topic name, for the fallback.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Open Fetch Gmail via IMAP. At the top of the command, replace <YOUR_EMAIL_ADDRESS>, <YOUR_APP_PASSWORD> and
   <YOUR_IMAP_HOST> with your own values. Keep the quotes.
3. Open Call Ollama Summary API and replace <YOUR_OLLAMA_MODEL> with the name of a model you have pulled. ollama list shows them.
4. Add a Telegram credential with your bot's token and select it in Send Telegram Summary. Replace <YOUR_TELEGRAM_CHAT_ID> in
   that node with your chat id.
5. In Send ntfy Summary, replace <YOUR_NTFY_TOPIC> in the address with your topic, and subscribe to the topic in the ntfy app.
6. Click Test workflow. Manual Trigger starts the run. It sends a message only when the model finds an urgent thread.
7. Check the workflow's timezone in its settings, since the schedule runs at 23:59 in it. Then save and press Publish at the top of
   the canvas (or Shift+P).

How a run goes

1. Manual Trigger or Schedule Trigger starts the run. The schedule fires every day at 23:59.
2. Fetch Gmail via IMAP runs a Python script. It logs in over SSL on port 993, opens the inbox and searches for mail since
   yesterday's date. It fetches each message with BODY.PEEK[], so nothing is marked as read, and keeps the ones from the last 24 hours.
3. The same script cuts each body to 2000 characters, groups the messages into threads and keeps up to 6 messages and 6000
   characters of text per thread. It prints one JSON object.
4. Parse JSON Output turns that output into an item. An empty output becomes an item with a note.
5. Build Summary Prompt mutes threads that match a list of patterns for promotions, job alerts, newsletters and similar. It
   marks urgent ones by another list. It prepares a fallback result of up to 5 urgent threads and 3 for info, and a context of
   up to 12 threads and 24000 characters.
6. Prepare Structured Prompt writes the system and user prompts. They ask for one JSON object with a summary, up to 5 urgent
   items and up to 3 items for info.
7. Call Ollama Summary API posts the prompts to /api/chat on your Ollama with temperature 0, a 32768-token context and at
   most 260 tokens of answer. It waits up to 90 seconds, and an error does not stop the run.
8. Parse Ollama Summary reads the JSON from the reply and trims each field. When the reply is not valid JSON, it uses the fallback result.
9. Has Urgent Items? lets the run go on only when at least one urgent item is left.
10. Normalize Summary Message writes the report: a summary, a text table of the urgent threads, up to 3 lines for info and a
    count of muted threads. It also makes an HTML version with the table in a pre block.
11. Fix Action Table Layout passes the item on unchanged.
12. Send Telegram Summary sends the HTML version. When it fails, Telegram Delivery Failed? sends the run to Send ntfy Summary,
    which posts the plain text with a title, a high priority and the tag mail.

Change the look

• The window: the cutoff and the search date in the script inside Fetch Gmail via IMAP.
• What counts as muted, urgent or for info: the pattern lists in Build Summary Prompt.
• The prompt and the answer limits: Prepare Structured Prompt.
• The model and its settings: the request body in Call Ollama Summary API.
• The look of the message: Normalize Summary Message. The table columns are 26, 18 and 30 characters wide.
• The time: Schedule Trigger.

Good to know

• Not run: Mindket made no call to an IMAP server, Ollama, Telegram or ntfy. It checked the Python script's syntax and ran the
  Code steps on its own machine with a sample thread and a sample model reply.
• The mail password sits in the command text. Anyone who can open the workflow can read it, so use an app password and keep the
  workflow to yourself.
• ntfy has no sign-up. Its documentation says the topic works like a password. The fallback text lists the subjects and
  senders of your urgent mail, so pick a topic nobody can guess.
• The mute patterns match word parts anywhere in the subject, the sender and the body, with no word borders. A real mail that
  contains "deal", "offer" or "sale" is muted before the model sees it. Edit the lists if that hurts.
• On a day with no urgent thread nothing is sent, not even a note that the day was quiet.
• The request sets think to false and asks for JSON output. Ollama's documentation describes think for thinking models.
• A 32768-token context needs more memory than Ollama's default, which its documentation puts at 4k on a machine with under 24 GiB
  of VRAM. The 260-token limit can cut a long answer short, and the workflow then falls back to its own lists.
• The message bodies go only to your own Ollama. The urgent subjects and senders go to Telegram, and to ntfy.sh if Telegram fails.
• A message can carry text meant to steer the model, so a mail can be listed as urgent or left out. Treat the report as a prompt to
  look, not a full check.