The Dialogue

I've been writing publicly since 2009. After 15 years and 640,000 words, I wanted to know: could a model answer questions the way I would?

What it does

The Dialogue is a retrieval-augmented generation model trained on my full corpus — essays, reporting, analysis, personal writing. You ask it a question. It finds the most relevant passages across everything I've published, then synthesizes an answer in my voice with direct citations back to the source pieces.

It's not a chatbot. It doesn't make things up. If I haven't written about something, it says so. Every answer is grounded in the actual text, with links to the original pieces so you can read them in full.

How it's built

The corpus is chunked, embedded, and stored in a Qdrant vector database. Queries run through a retrieval pipeline that uses both dense (OpenAI) and sparse (BM25 via Cohere) retrieval — hybrid search surfaces better results than either alone. Claude handles the synthesis step, stitching retrieved passages into a coherent answer that reads like I wrote it.

The API is a FastAPI service deployed on Railway. The front end is intentionally minimal — a text input, an answer, and citations. The writing is the product.

What I learned

You learn things about your own prose that only a machine can surface. I repeat myself more than I thought. Certain themes — the gap between intention and execution, the compounding value of consistent work — show up across wildly different topics. The model knows my tics better than I do.