Dr. ROCM — Medical Image Triage Assistant

Streamlit app that takes an uploaded image (X-ray, clinical photo, or prescription) and returns a structured triage read using a vision model hosted on Cloudflare Workers AI, called through Cloudflare AI Gateway (for logging, caching, and rate limiting). No model runs locally, so it deploys cleanly on Streamlit Community Cloud's free tier.

⚠️ Triage assistance only — not a diagnosis.

Getting Cloudflare credentials

1. Account ID — visible on the right sidebar of any page in the Cloudflare dashboard, or run wrangler whoami.
2. API token — My Profile → API Tokens → Create Token → grant it the Workers AI: Read permission (this is required even for third-party/gateway calls, per Cloudflare's AI REST API).
3. Gateway ID — go to AI → AI Gateway in the dashboard. A gateway named default is created automatically the first time you use it, or you can create a named one for its own logs/analytics.

Run locally

pip install -r requirements.txt
export CLOUDFLARE_ACCOUNT_ID=your_account_id
export CLOUDFLARE_API_TOKEN=your_api_token
export CLOUDFLARE_GATEWAY_ID=default
streamlit run app.py

Deploy on Streamlit Community Cloud

1. Push this folder to a GitHub repo.

2. On share.streamlit.io, create a new app pointing at app.py in that repo.

3. In the app's Settings → Secrets, add:

   CLOUDFLARE_ACCOUNT_ID = "your_account_id"
   CLOUDFLARE_API_TOKEN = "your_api_token"
   CLOUDFLARE_GATEWAY_ID = "default"

4. Deploy. If secrets aren't set, the sidebar falls back to plain (unmasked account ID, masked token) input fields so you can paste credentials in at runtime instead — handy for quick demos.

Credentials & the sidebar

• If CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN are set (env vars or Streamlit secrets), the sidebar just shows "Cloudflare credentials loaded from server config ✓" — the actual values are never echoed into a visible or editable widget, so they can't be read off the screen by anyone viewing the app.
• A collapsed "Use different credentials for this session" expander lets you override them for one session without touching the configured secrets; it starts blank.
• If no credentials are configured at all, the normal input fields appear (account ID plain text, API token masked).

Choosing a model

The sidebar lets you switch between three Workers AI vision models without touching code:

• Llama 3.2 11B Vision (default) — Meta's dedicated vision-instruct model, a solid general choice for image reasoning and OCR-style extraction (useful for the prescription-text case).
• Llama 4 Scout 17B — natively multimodal MoE model, generally stronger reasoning, a bit more expensive per token.
• Mistral Small 3.1 24B — adds vision on top of Mistral's small dense model.

Swap the VISION_MODELS dict in app.py to add others from Cloudflare's Workers AI model catalog (anything tagged "Vision").

How the triage flow works

1. Upload an image and (optionally) add patient context, then run the analysis. The result renders as a report card: image type, triage label, summary, findings, prescription text (if any), and up to three follow-up questions.
2. Follow-up questions appear as buttons, not a free-text chat. Click one to open a text box, write your answer, and save it — the question gets a ✅ and you can open and answer as many others as you like first.
3. There's also an optional "ask something else" field for anything not covered by the suggested questions.
4. When ready, click 📤 Send answers to model — all saved Q/A pairs (plus your custom message, if any) are sent together in a single message, with the image still in context.
5. The model's reply comes back in the same structured format as the first analysis, so it renders as another report card — with its own fresh set of follow-up question buttons — right below the previous one. Repeat steps 2–5 until you're satisfied with the read.
6. Replies are parsed with a regex-based field extractor (not a strict newline split), since the model doesn't always put each field on its own line. If parsing ever fails to find a recognizable triage_label, the app falls back to showing the raw reply text inside the card rather than crashing.

Notes on this version

• No local model. The original build loaded Qwen2-VL-2B locally via transformers/torch on a Hugging Face ZeroGPU Space. That doesn't run on Streamlit Cloud (no GPU, ~1GB RAM), so image analysis now calls a hosted model through Cloudflare AI Gateway instead.
• No ChromaDB / sentence-transformers. Vector memory added a second heavy torch-based dependency and wrote to local disk, which Streamlit Cloud wipes on every redeploy anyway. Triage history is now kept in the session (sidebar) and best-effort logged to data/triage_log.json on the running instance — treat that file as ephemeral, not a database.
• If you outgrow the free tier's storage limits and want real persistent memory, swap log_record() for a call to a hosted store (Postgres, Supabase, a hosted vector DB, etc.) — the rest of the app doesn't need to change.