README.md

metadata

title: Ai Econsult Demo
emoji: 🚀
colorFrom: red
colorTo: red
sdk: streamlit
app_port: 8501
tags:
  - streamlit
pinned: false
short_description: Streamlit template space
sdk_version: 1.51.0

AI‑Native E‑Consult Prototype (V1)

This Space hosts a Streamlit app for an AI‑native e‑consult workflow (PCP ↔ Specialist).

• UI: Streamlit
• AI: MedGemma‑27B‑text + optional RAG grounding
• Storage: Hugging Face Spaces persistent volume (/data/econsult/...)

Modes available

• 🧠 Guideline‑grounded (RAG): Injects top guideline snippets into the prompt. Expects inline bracket citations [1]..[k] plus a compact reference list.
• ✨ AI‑only (post‑hoc): No in‑prompt evidence. Performs post‑hoc claim→guideline mapping and returns a compact reference list.

The previous Hybrid (AI → RAG) option has been removed (redundant for this prototype).

Hardware

• Primary: 1× L40s (works well for google/medgemma-27b-text-it)
• Optional upgrade: A100 (large)
• Budget fallback for demos: L4 with google/medgemma-4b-it

CPU fallback is supported for resilience only and will cap tokens for responsiveness. You can disable CPU generation with ALLOW_CPU_GENERATION=false.

Setup

1. Add your Hugging Face token in Settings → Variables as HF_TOKEN (required to pull gated models).
2. (Optional) Set environment variables:
   • ALLOW_CPU_GENERATION=true|false (default: true)
   • FORCE_CPU_LLM=1 to force CPU fallback to the 4B model
   • MODEL_FALLBACK_ID=google/medgemma-4b-it to override the fallback id
3. Upload guideline PDFs on the “Step 1 – RAG Corpus Prep” page. Filenames should follow:
   [specialty]_guideline_[condition]_year.pdf

What changed in this revision

• Removed AI triage across UI and storage. The PCP must choose a Consult specialty.
• Removed the Hybrid button on the Specialist tab.
• Added a demo‑only / de‑identified banner to all pages.
• Preserved citations from AI‑only generation (post‑hoc mapping).
• Improved guideline filename parsing to your taxonomy ([specialty]_guideline_...).
• Added a per‑specialty count widget on the RAG page.
• Cleaned up docs for native Spaces runtime. Removed the Dockerfile (not used).

Notes

• This prototype is not for clinical use. Use only de‑identified/mock data in demos.
• After you confirm stability on Spaces, we can pin dependency minor versions for reproducibility.