FileLoader with Postgres (pgvector) does not remove old embeddings — `documents` table grows indefinitely with each new upload

[jeme95]

Describe the bug

When using Flowise in a Docker environment with a Document Store and a FileLoader, connected to Postgres with pgvector, embeddings from old files remain in the database even after new files are uploaded.
Flowise always appends new embeddings to the documents table without removing older entries that are no longer needed, causing the table to grow indefinitely. Flowise becomes noticeably slower — both the "Process" step and the "Upsert" step take significantly longer, making the system sluggish over time

To Reproduce

Steps to Reproduce:

1. Run Flowise in a Docker container with a Postgres (pgvector) database connected as Document Store.

2. Create a FileLoader node and upload one or more files.

3. Check the documents table in the Postgres database — embeddings are stored as expected.

4. Remove the old files from the FileLoader input and upload completely new files.

5. Check the documents table again — the old embeddings are still present, and new embeddings have been appended instead of removing outdated ones.

Expected behavior

old embeddings should be deleted from the documents table, because they are no longer part of the active Document Store

Screenshots

Flow

No response

Use Method

None

Flowise Version

3.0.5

Operating System

Windows

Browser

Chrome

Additional context

No response

[hemisalo]

Possible duplicate of #3570 and #4152 .

This is an issue when Postgres is used as Vector Store. Not sure if using Postgres as RecordManager affects the issue or not.

There's a PR #4808 That seems to be a fix for this. I have temporarily "fixed" the issue by setting up and using Supabase as the Vector Store.

@HenryHengZJ This issue has been around since Nov 2024, might be worth a look if the discussion in #3570 and the PR #4808 are of any help/are a good solution.

[onestardao]

this matches no.7 – memory breaks across sessions from our 16-problem list, except here it’s in the persistence layer — embeddings are never fully removed, so the documents table just keeps growing and old vectors still get surfaced.

mit licensed project, cold start to ~600 stars in 60 days.
if you want the cleanup + retention guard patterns we use to fix this class, let me know and i can share.