Gmail to Vector Embeddings with PGVector and Ollama

Who is this for?

Everyone! Did you dream of asking an AI "what hotel did I stay in for holidays last summer?" or "what were my marks last semester like?".

Dream no more, as vector similarity searches and this workflow are the foundations to make it possible (as long as the information appears in your e-mails 😅).

100% local

This workflow is designed to use locally-hosted open source. Ollama as LLM provider, nomic-embed-text as the embeddings model, and pgvector as the vector database engine, on top of Postgres.

But.. how?!

Firstly, specify the date you created your Gmail account on, then manually run the workflow in order to bulk read all your e-mail in monthly batches. Your database is now populated!

Now it's the task for other workflows to query the vector database.

Activate the workflow so that new e-mail is continuously added by the Gmail Trigger upon receiving it.

Structured AND Vectorized

This workflow stores your e-mail activity in two ways:

• In a structured table
• In a vector embeddings table

And the information in both of them can be correlated by Gmail's messages id, which is stored in the vectors table as metadata property emails_metadata.id.

That way consumers can benefit from both worlds! ✨ Vector similarity searches enable semantic searches, while structured queries can retrieve more factual data like the message id, its date or who it came from.

Other useful templates

My template Chat with Your Email History using Telegram, Mistral and Pgvector for RAG is a ready-made solution to consume this workflow.

You may also pair this workflow with my other template to Email Assistant: Convert Natural Language to SQL Queries with Phi4-mini and PostgreSQL and you'll enable RAG workflows that use both structured and vectorized databases.

Customizations

I suppose the e-mail provider could be changed, but then you'd have to identify an alternative id field. Message-ID would be a more standard option.

There are a few opinionated choices as to what metadata to store, but those shouldn't need adjustments.