Enhance Email Indexation Metadata for RAG Source Recognition

[AlaeMghirbi]

Why

Currently, the Cozy front-end can only display Cozy Files as sources for the AI assistant when generating responses using RAG .

However, when the AI assistant uses an email as a source, the front-end doesn’t have enough information about the email to recognize or display it properly.

To improve this, we want to send the necessary email information during the indexation phase so that later the Cozy front-end will be able to:

• Recognize indexed emails as RAG sources.
• Display them properly.
• Provide a way to redirect users to the original email mentioned in the source.

---

What

We focus on sending the required metadata to the Cozy indexation database.

We need to send detailed information about the email, including:

• Date
• Subject
• Sender
• Recipient
• **Email preview **

We also need to provide enough metadata to allow Cozy Mail to later build a redirection link to the corresponding email.
For this, we must include the emailId in the indexed metadata.

---

How

When indexing an email, the TMail backend currently sends basic metadata such as the date and document type through the following method:

computeMetaData located in tmail-backend/tmail-third-party/ai-bot/src/main/java/com/linagora/tmail/mailet/rag/RagListener.java

We need to update this method to include the following additional metadata fields:

• email subject
• email sender
• email recipient
• email preview

To enable reliable redirection to the corresponding email in the Cozy front-end, we must also:

• Include the emailId in the metadata or even replace the current usage of messageId (currently used as the name of the indexed email) with emailId, since messageId will no longer be required later.

[chibenwa]

Question: can we name the metadata passed to the rag? The format (string? Unix timestamp? Mailbox format for mail address?), are they optional, etc...

[AlaeMghirbi]

Question: can we name the metadata passed to the rag? The format (string? Unix timestamp? Mailbox format for mail address?), are they optional, etc...

I asked Paul about the specifications for the metadata to ensure consistency with what is implemented for Cozy Files, and he showed me how emails are currently presented.

In the picture below, a file is displayed like this in the assistant (here, it is actually an email converted into a .md file):

---

Final Display Adjustments

The final display should remain roughly the same, with the following adjustments:

• The icon should be the Twake Mail icon
  → cozy-twakemail/public/icon.svg
• The title should display the email subject
• The secondary title (currently the file path) could display one of the following:
  • The email sender, or
  • The email directory (with a default “mailbox” if none exists), or
  • The email preview

---

There are no optional fields required to guarantee the display of the email, but we handle empty values in the Cozy backend to prevent errors.

For the field names, I checked the Cozy backend, and they are retrieving this information directly from the RAG response like this:

        doctype, _ := src["doctype"].(string)
		file_id, _ := src["file_id"].(string)
		file_name, _ := src["filename"].(string)
		page, _ := src["page"].(float64)
		file_url, _ := src["file_url"].(string)
		chunk_url, _ := src["chunk_url"].(string)

So, I think it would be better to keep the names doctype and filename.
In our case, filename could be a combination of date + subject to match the expected display format.

[AlaeMghirbi]

Question: can we name the metadata passed to the rag? The format (string? Unix timestamp? Mailbox format for mail address?), are they optional, etc...

Paul also added some comments in the notion spec : https://www.notion.so/linagora/Email-Metadata-Specification-for-RAG-Indexing-28e62718bad180efa1a7eadc08f66e9d