What started a few years ago as a RAG exercise - pgvector, ChromaDB, LangChain, a few Python scripts - is now Chatty, the eleventh app in the MyOwnLifeOS suite. It is a utility I use daily to ask questions about my own documents: uploaded files, Google Drive contents, notes, spreadsheets, PDFs, and legacy Word files. Every answer comes from a local Ollama server. No document text leaves the network.
The point is utility. Chatty answers two kinds of questions. Catalog questions like "list all my PDFs" or "who shared the Q3 report" are answered from stored metadata. Content questions are turned into embeddings, matched against the indexed chunks, and fed to the model with the last few turns of conversation. The response streams back, and a sources footer names the exact documents that answered.
Local inference, not API calls. Embedding and chat both run on the same Ollama box. Models are pre-warmed at startup and kept resident so context survives between turns. Even the stop button is scoped to the individual browser session, so one user stopping a response cannot cancel another user's stream.
Two backends, one app. Chatty ships as either a pooled Postgres + pgvector deployment or a local ChromaDB store. Both share the same Gradio UI and the same answer flow. The vector store is a deployment choice, not a different product.
Feedback becomes training data. Every answer can be thumbs-up or thumbs-down. A dislike shows up in the Library tab and can be turned into a correction, which is indexed like any other document and competes for retrieval the next time the same topic comes up. The sources footer plus the correction loop are what make the app accountable for what it actually said.
Locked by default. Every route except /health sits behind session auth. The email allowlist fails closed. Google OAuth grants identity and read-only Drive access in one consent, and the refresh token is only saved for an authorized account.
The full architecture is public at chatty.ifthenelse.net/architecture: the auth gate, the shared UI shell, the interchangeable backends, the Drive sync path, and where Chatty fits in the suite.
Chatty is not a wrapper around a cloud API. It is a small in-house software factory for a single user: ingest, embed, retrieve, answer, correct, repeat.
It is live at chatty.ifthenelse.net for anyone curious enough to look, though the instance itself is invite-only. If you have your own Ollama server and a pile of documents you would rather not paste into a third-party model, the pattern is worth copying.