AI For Dev

summary-cole-medin-local-supabase-rag

2 min read

Run Supabase 100% LOCALLY for Your AI Agents

Source: YouTube — Cole Medin, published 2025-02-16 Tools: Supabase (self-hosted), n8n, Ollama, Open WebUI, Postgres, PGVector, Qdrant, Flowise

Summary

Cole Medin revamps his local AI package (coleam00/local-ai-packaged) to add fully self-hosted Supabase alongside the existing n8n / Ollama / Open WebUI / Qdrant / Flowise stack. The new piece is wiring Supabase + PGVector as the vector store for a local RAG agent (replacing or complementing Qdrant), with Ollama’s nomic-embed-text (768-dim) as the embedding model. Demo: crawl a local file → insert chunks into Supabase vectors → query via an n8n agent → grounded answer. Includes troubleshooting notes for the Supabase Docker stack (multiple containers: studio, Kong, auth, realtime, storage, functions) and persistent volume / host path sharing config so n8n can read and write the same data.

The local-AI-packaged stack (post-update)

n8n (workflow / agent)
  ├── Postgres (chat history, agent state)
  ├── Qdrant (vector store, optional)
  ├── Supabase (self-hosted: studio + Kong + auth + storage + Postgres + PGVector)
  └── Ollama (LLM + embeddings)
Open WebUI (chat UI)
Flowise (visual agent builder, optional)

All in a single docker-compose orchestrated by coleam00/local-ai-packaged.

Key facts

  • Repo: github.com/coleam00/local-ai-packaged
  • Supabase containers: studio, Kong (gateway), auth (GoTrue), realtime, storage, functions, Postgres
  • Vector store: PGVector inside Supabase Postgres — replaces Qdrant for the RAG agent path (Qdrant remains optional)
  • Embeddings: Ollama nomic-embed-text (768-dim)
  • Persistence: explicit volume mounts; host-path sharing so n8n containers can read files placed on the host

Why it matters

This is the wiki’s canonical local-AI stack reference — every component is self-hosted, no cloud calls required, and the components are the same ones that show up across the rest of the wiki (n8n, supabase, ollama, open-webui, pgvector). The PGVector-as-vector-store choice is also the load-bearing connection back to the OpenBrain thread: OpenBrain is “Postgres + PGVector + MCP” as agent memory, and Cole’s stack is the practical recipe that makes both that and traditional RAG run on the same database.

See Also

Graph View

Backlinks