Sunday Coffee & Code: More OpenRAG + a UI layer for my agents

Today was a two-track build day: getting deeper into OpenRAG, and putting a simple UI in front of my agent stack. 1) OpenRAG progress I got IBM’s OpenRAG up and running and configured with Ollama + Granite. The document ingestion is giving me trouble (500 error), so that’s next week’s problem. I also spent time looking through the LangFlow canvas at all the flows and was reminded how productive that interface is when you’re iterating quickly. A couple of practical wrinkles along the way: ● I moved Ollama to the public interface so LangFlow could see it and call the models (we are running completely offline). ● I switched to CPU mode in the TUI to get rid of the noisy GPU messages (which should be addressed by a recent PR). 2) A UI for the agents I switched gears and started putting a UI on the front of the agents. Claude Code did an amazing job again. Credit to René Fournier for pointing me to Estimate Hawk (https://estimatehawk.com/ - great interface and a very similar concept). My PoC UI is based on that approach - Claude Code built the PoC UI from a snippet of the web page and nailed it (will be redesigned for future releases). Architecture-wise, I made a few key moves: ● Separated the UI from the upload/status/download functionality to improve portability. ● Built a small daemon to handle uploads, status, and download. ● From a security standpoint, I completely decoupled the multi-agent orchestrator and agent stack from the FE/BE UI. The orchestrator will run autonomously on a trigger basis, which also makes it easier to integrate other front-ends later (i.e. very loosely coupled). There’s still a potential prompt injection risk via uploaded documents - but mitigating that is part of a future “security agent” in the pipeline.