The model is not given the repos. Every question runs one loop: find the few chunks that matter, paste only those into the prompt, answer with citations.
The model runs three times here. Every other stage is SQL or arithmetic. Click a stage for what it does and the numbers it just produced.
Each box is one ingest step, with the number it actually produced. Click it for the mechanism.
Every box below is a real stage. Numbers appear when a stage finishes, and they are the numbers that stage produced. Click a box to open it, and Next inside the panel walks the stages in order.
The model decides twice in this graph: whether to retry, and the words of the answer. Which chunks enter the prompt is decided by arithmetic over the two searches and the graph, never by the model.
Only real runs appear here. Nothing on this page is mocked.
Fifteen fixed questions with known answers, run through the real pipeline. Retrieval is scored on its own: if the right file never came back, no prompt can fix the answer.
Themes next to this pipeline rather than inside it.