Implementation Overview

This project is a Next.js app that generates and evaluates SOAP notes from clinical transcripts.

• Frontend (App Router)

  • app/page.tsx – Upload transcript, choose model, and start SOAP note generation.
  • app/results/page.tsx – Shows uploaded transcript and generated SOAP note side by side, lets you upload a reference SOAP note, and triggers evaluation.
  • app/analysis/page.tsx – Displays evaluation analysis with a radar chart (ROUGE‑1, BLEU‑1, combined score), model details, token counts (input/output), and the three texts: generated SOAP, reference SOAP, and transcript.
  • State is persisted across pages using localStorage keys such as soap_transcript, soap_reference, soap_results, and soap_metrics.

• Generation API

  • app/api/generate/route.ts (POST /api/generate).
  • Accepts a JSON payload with transcript and selected model (OpenAI or Gemini).
  • Builds a structured prompt that asks the model to produce a complete SOAP note in four sections (S – Subjective, O – Objective, A – Assessment, **P – Plan`).
  • Calls either:
    • Gemini via https://generativelanguage.googleapis.com/v1/models/{model}:generateContent, using GOOGLE_API_KEY, or
    • OpenAI Chat Completions via https://api.openai.com/v1/chat/completions, using OPENAI_API_KEY.
  • Returns { message, modelUsed }, where message is the generated SOAP note (with **...** used for bold headings).

• Evaluation API

  • app/api/evaluate/route.ts (POST /api/evaluate).
  • Accepts a JSON payload with reference (reference SOAP note) and candidates (generated notes).
  • Uses simple tokenization to compute:
    • ROUGE‑1 (overlap‑based F1 score),
    • BLEU‑1 (unigram precision with brevity penalty),
    • Combined score = average of ROUGE‑1 and BLEU‑1.
  • Also computes and returns per‑candidate token counts:
    • tokenInfo.referenceTokens (input tokens),
    • tokenInfo.candidateTokens (output tokens).
  • Response shape: { metrics: [{ id, label, rouge1, bleu1, combined, tokenInfo }] }.

• Downloadable Results

  • The analysis page can export a JSON file summarizing the run: timestamps, model/provider, transcript, reference SOAP, generated SOAP, evaluation scores, and token information.

Running on Vercel

This app is designed to run as a standard Next.js App Router project on Vercel.

1. Environment variables

   • Configure these in your Vercel project settings (Dashboard → Project → Settings → Environment Variables):
     • OPENAI_API_KEY – for OpenAI models (if you want OpenAI generation).
     • GOOGLE_API_KEY – for Gemini models (if you want Gemini generation).

2. Deploying

   • Push this repository to GitHub, GitLab, or Bitbucket.
   • In Vercel, create a new project and import this repository.
   • Vercel will detect Next.js automatically and use the default build settings:
     • Build command: npm run build
     • Install command: npm install
     • Output: Next.js app (no manual output path configuration needed).

3. Runtime behavior on Vercel

   • API routes app/api/generate/route.ts and app/api/evaluate/route.ts run as Vercel Serverless Functions.
   • Frontend pages (/, /results, /analysis) are served as Next.js routes and use the same logic as in local development.
   • As long as the environment variables are set correctly, the deployed app on Vercel will be able to generate SOAP notes, evaluate them, and show/download the analysis without additional configuration.