I've been working more on my local homelab LLM project. It runs gpt-oss:20b as its base model in conjunction with a RAG system (BM25+embeddings) accessed with a custom web UI. Nothing groundbreaking there. What's exciting for the geek in me is how the system uses what I call policy-driven RAG. Spoiler, no LLM is truly rule based. What my system does is uses a file called other_topics.txt to determine if model output comes solely from RAG data (the default), from general gpt-oss:20b model data (aka other topics), or a hybrid mode. Does it scale to larger deployments? I have no idea. But it's very homelab geek cool. #opensource #localllm #pangeahillsai #openai #gpt-oss #homelab

@benjamineskola Yes, it's an LLM, they operate in a world of probability. The decision for the model to evaluate the prompt from only RAG Corpus vs Model Corpus vs a hybrid is "controlled" by a RAG document that spells out which route the model should take. So questions about my homelab route to strictly RAG data. Questions about a US state route to the Model's general knowledge. Queries about politics would route to the RAG stating there is no data in RAG documentation, because politics is not listed as a source for local LLM knowledge. It's not 100% bullet proof, but it is extremely accurate in letting me currate the route I want the model to take. I'm a soon to be retired firefighter, not an LLM expert by any means. I'm just having fun in my homelab and letting other people know what I'm learning. I hope that I learn a lot from others with more experience than myself. Thanks.

@benjamineskola Totally get what you’re saying — but that’s actually not how my setup works.

I’m not claiming the LLM “understands” rules in a human sense. What does enforce policy is the system around it: the router layer, the Flask server, the tool sandbox, and the guards that define what the model is allowed to call or see.

The LLM is just the reasoning engine.
The policy compliance comes from the scaffolding.

Same way a browser doesn’t “know” web security but still follows CSP, sandboxing, and permissions because the runtime enforces it. My local setup works the same way: external rules, internal tools, and capability boundaries.

“So yes — it’s probabilistic, but it only operates inside the fences I build. I’m just glad I found a technique that’s easy to configure and maintain at the homelab level. I keep BM25 and vector weights very low so whether it uses homelab data or general knowledge is strictly driven by the indexed, chunked RAG corpus.”

I call my homelab bitfrost.lan. It's a play on bifrost in Norse methology about the rainbow bridge. So when I started my local LLM project, I naturally called it bitfrost. AI. Unfortunately, that is a real site, not related to me at all. I'm now calling my project PangeaHills. I'm retiring soon to the North Georgia foothills. The Appalachian Mountains are the ancient remants created when Pangea was formed. Thus, PangeaHills. Happy Geeking. #homelab #geek #pangeahillsAI

Proof-of-concept:
I ran Glances inside Apple’s new native container runtime on macOS 26 — I call it Gamur (Icelandic for “container”).
Even cooler: Gamur Glances monitors Docker Glances.
Apple nailed this.

#macOS26 #AppleSilicon #containers #homelab #virtualization #DevOps #Glances

Another RAG-to-Riches story.
…ok, not really — I’m just an average homelab goblin who accidentally reinvented a thing the real AI folks probably solved 3 years ago. 😂

But I call it policy-driven RAG:
where rules, not model weights, decide when my local LLM answers from the doc corpus and when it’s allowed to use general knowledge.

Basically:
“Stay in your lane unless I say otherwise.”
RAG for homelab stuff
LLM for the wider world
Never phones home

Might be old hat to the pros, but it absolutely blew my geek mind that it actually works.

Just a homelab nerd living his best offline-AI life.

#homelab #LocalAI #RAG #SelfHosted #PangeaHillsAI #LLM #NerdLife #ServerGoblin #ComputersAreFun

Built something fun in the lab: PangeaHills.ai, my own locally-hosted, policy-driven RAG + LLM stack.
Completely offline, totally self-contained, powered by a bunch of noisy equipment pretending to be a cloud. 😄

The neat part? It’s rule-driven, not weight-driven:
• Homelab questions must stay inside the RAG universe.
• General topics only switch to model knowledge when my routing rules explicitly allow it.
• The LLM never “guesses” when to leave RAG — it follows policies, not vibes.

Feels like having an AI that actually stays in its lane because you built the lane lines yourself. 🚧🤖

#homelab #selfhosted #LLM #RAG #PolicyDrivenAI #SelfHostedAI #HomeLabLife #BSD #Linux #PangeaHillsAI #nerdlife