All projects
RAG / Vector Databases / Data Engineering

AI Record Manager & Vector Database Sync

A workflow that synchronises external content with a Pinecone vector database while maintaining a record manager in MySQL, so embeddings reflect the current state of the source rather than drifting out of date.

Year
2025
Engagement
Automation & data engineering
Focus
Data Engineering, Generative AI, Automation
Status
Open source
AI Record Manager & Vector Database Sync — architecture overview

01 — The problem

A RAG system is only as current as its index. Source content changes, gets deleted, gets re-added — and a vector store with no memory of what it already ingested will happily serve stale or duplicated answers.

02 — Our solution

A record manager in MySQL tracks what has been ingested and in what state, so each sync run knows what is new, what changed and what should be removed from Pinecone. The vector store becomes a projection of the source rather than an append-only pile.

Architecture

How it works

The pipeline, step by step — from the first input to the final output.

  1. 01

    Source read

    External content is read on a schedule or trigger.

  2. 02

    Change detection

    The MySQL record manager is compared against the source to classify each item as new, changed, unchanged or removed.

  3. 03

    Embedding

    New and changed content is chunked and embedded.

  4. 04

    Pinecone upsert

    Vectors are upserted, and vectors for removed content are deleted.

  5. 05

    Record update

    The record manager is updated so the next run starts from an accurate baseline.

Features

What we built

MySQL record manager as the source of ingestion truth
New / changed / unchanged / removed classification per run
Incremental embedding instead of full re-index
Deletion handling so removed content stops being retrievable
Impact

The outcomes

01

An index that stays consistent with its source

02

Incremental syncs rather than full rebuilds

03

A reusable pattern for any RAG ingestion pipeline

Engineering notes

Challenges we solved

Deletes are the hard part

Adding vectors is easy; knowing what to remove requires state. The record manager exists precisely to answer that.

Avoiding wasteful re-embedding

Re-embedding everything is simple and expensive. Change detection keeps each run proportional to what actually moved.

Technology stack

n8nPineconeMySQLEmbeddingsRAGVector Search

Want results like these?

Book a free 30-minute discovery call. We'll map out your project, recommend the right stack and send a fixed quote within 48 hours.