TrinetraLabs/Placebo_AI

<div align="center">

🧠 Placebo AI

Medical Intelligence Platform — Serverless RAG at LPU Speed

    

Instant, medically-grounded answers backed by 1.3M+ medical text vectors — with zero local disk footprint.

Live Demo · Architecture · Setup · Deployment

</div>

📖 Overview

Placebo AI is a cloud-native medical intelligence platform that combines retrieval-augmented generation (RAG) with a fully serverless data layer. Rather than shipping a multi-gigabyte vector index alongside the application, Placebo AI offloads all heavy assets — embeddings, document text, and images — to managed cloud services, leaving the application itself lightweight, stateless, and trivially deployable to any container platform.

The result is a system that can answer nuanced medical questions with grounded, source-backed context, streamed back to the user in near real-time thanks to Groq's LPU (Language Processing Unit) inference engine.

Disclaimer: Placebo AI is an informational and educational tool. It is not a substitute for professional medical diagnosis, advice, or treatment. Always consult a qualified healthcare provider for medical concerns.

🏗️ Architecture

Placebo AI's architecture was deliberately redesigned to eliminate the operational fragility of local, disk-resident datasets. The migration removed a 17GB vector store and an 88GB image corpus from the application's runtime footprint entirely, replacing them with serverless equivalents that scale independently of the app container.

flowchart LR
    U[User Browser] -->|HTTPS| FE[Frontend<br/>Glassmorphic UI]
    FE -->|JWT Auth| AUTH[Supabase Auth]
    FE -->|Query| API[FastAPI Backend]

    API -->|Keyword-Augmented<br/>Retrieval| ASTRA[(Astra DB<br/>Serverless Vector Search)]
    API -->|Asset Fetch| HFDS[(HuggingFace Datasets<br/>Docs & Images)]
    API -->|Grounded Prompt| GROQ[Groq LPU<br/>Llama-3.1-8b-instant]

    ASTRA -->|Top-K Context| API
    HFDS -->|Media/Refs| API
    GROQ -->|Streamed Tokens| API
    API -->|SSE Stream| FE

Design Principles

🧩 Tech Stack

⚡ Key Features

• —Zero Local Footprint — The former 17GB vector database and 88GB image archive have been fully migrated off-disk to Astra DB and HuggingFace Datasets respectively. The running application requires negligible local storage.
• —Lightning-Fast Inference — Powered by Groq's LPU architecture, responses stream token-by-token with minimal latency compared to traditional GPU-based inference.
• —Keyword-Augmented RAG Retrieval — A custom retriever layer blends semantic vector search with keyword-based re-ranking to improve relevance on medical terminology and abbreviations.
• —Strict XML Prompt Boundaries — All retrieved context is wrapped in explicit XML tags before being passed to the LLM, reducing prompt injection risk and improving grounding fidelity.
• —Secure Authentication — Supabase-issued JWTs are verified server-side on every protected request, ensuring session integrity without maintaining server-side session state.
• —Accessibility & Performance — The frontend achieves perfect 100/100 Accessibility and SEO scores on Google Lighthouse, with a responsive, glassmorphic clinical interface.
• —Platform-Agnostic Deployment — Being fully serverless on the data side, the same Docker image can be deployed to HuggingFace Spaces, Render, Railway, Vercel, or any container-compatible host without modification.

🔌 API Overview

Update endpoint paths/names below to match your actual FastAPI routes — placeholders shown for structure.

🛠️ Installation & Setup

Prerequisites

• —Python 3.10+
• —A Groq API key
• —A DataStax Astra DB instance (serverless vector-enabled)
• —A Supabase project (for authentication)

1. Clone the repository

git clone https://github.com/labstrinetra/placebo.ai.git
cd placebo.ai

2. Install dependencies

pip install -r requirements.txt

3. Configure environment variables

Create a .env file in the project root:

GROQ_API_KEY=your_groq_api_key
ASTRA_DB_API_ENDPOINT=your_astra_endpoint
ASTRA_DB_APPLICATION_TOKEN=your_astra_token
SUPABASE_URL=your_supabase_url
SUPABASE_ANON_KEY=your_supabase_key

4. Run the application

python src/app.py

The application will be available at http://localhost:8000.

📦 Deployment

Placebo AI is live in production on HuggingFace Spaces, running in a custom Docker container:

🔗 Live URL: https://trinetralabs-placeboai.hf.space/

Because all heavy data lives off-container, the exact same repository can be deployed to any of the following without disk or memory tuning:

• —Render — Docker web service, no persistent disk required
• —Railway — one-click Docker deploy from GitHub
• —Vercel — via Docker-compatible serverless functions (API routes)
• —HuggingFace Spaces — current production host (Docker SDK)

Deployment Checklist

• —[ ] All environment variables set in the host platform's secrets manager (never commit .env)
• —[ ] app_port matches the platform's expected port (7860 for HF Spaces, configurable elsewhere)
• —[ ] Astra DB and Supabase projects are on production-tier plans if expecting sustained traffic
• —[ ] Groq API rate limits reviewed for expected concurrent load

🔒 Security Notes

• —JWTs issued by Supabase are verified server-side on every authenticated request; no session state is persisted in the application.
• —API keys and database tokens are read exclusively from environment variables — never hardcoded or logged.
• —Retrieved context is sandboxed within explicit XML boundaries before being passed to the LLM, limiting the blast radius of prompt-injection attempts embedded in source documents.

🗺️ Roadmap

• —[ ] Multi-turn conversational memory with session-scoped context windows
• —[ ] Source citation UI (inline links to retrieved documents/images)
• —[ ] Admin dashboard for corpus management and query analytics
• —[ ] Rate limiting and usage quotas per authenticated user

🤝 Contributing

Contributions, issues, and feature requests are welcome. Please open an issue first to discuss significant changes before submitting a pull request.

📄 License

Distributed under the MIT License. See LICENSE for details.

<div align="center">

Developed by Trinetra Labs**

</div>