Health Docs

Every model call goes through a service that resolves to Ollama or to a hosted API from one environment variable. Writing that seam cost an afternoon at a moment when the hosted model was the only one good enough to build against. It is what let the project be tested seriously and then moved home without touching a controller. The version that runs is the local one.

Nothing in the code knows what a medical folder tree looks like. On each document the model receives the tree as it stands, with document counts, and answers use_existing or create_new with a path. A fixed taxonomy would have been simpler and would have been wrong by the third unusual document, which in a medical history is roughly every document. The bill for that choice is in Would change, below.

The hosted model returns the JSON you asked for. A local 8B model returns the JSON plus a sentence explaining how helpful it has been, sometimes in a fenced block, sometimes with a trailing comma or a stray comment. Rather than retry until it complies, which burns seconds on a single GPU, the local path takes the first brace to the last and repairs the usual damage before parsing. It is not elegant and it is why the local provider became usable at all.

const firstBrace = cleanedResponse.indexOf('{');
const lastBrace = cleanedResponse.lastIndexOf('}');

if (firstBrace !== -1 && lastBrace !== -1 && lastBrace > firstBrace) {
  cleanedResponse = cleanedResponse.substring(firstBrace, lastBrace + 1);
}

// Trailing commas, // comments, single quoted keys
jsonString = jsonString.replace(/,(\s*[}\]])/g, '$1');

A hand-maintained medical summary is wrong within a month. Here the model reads the whole current report and the new document, and returns the whole report again, so the documents stay the thing that has to be right. Asking for a diff would be cheaper and would have meant merging model output into a document by hand, which is exactly the class of bug you do not want in a medical record.