Confidence

A RAG-powered skincare intelligence system - selfie to personalised routine in under 60 seconds.

Live: https://confidence-two.vercel.app | API: https://confidence-api-59597652459.us-central1.run.app/health

---

What is RAG and why it matters here

Without RAG, a language model generates skincare recommendations from its training data — generic, unverifiable, hallucinated product names and ingredients.

Confidence uses Retrieval-Augmented Generation. The LLM never invents products. Every recommendation traces back to a real row in a vetted product database.

---

How RAG works in Confidence

flowchart LR
    subgraph INDEX ["INDEX PHASE — run once"]
        A[100 skincare products] -->|embed| B[Voyage AI\nvoyage-4-lite]
        B -->|store 1024-dim vector| C[(Supabase\npgvector)]
    end

    subgraph QUERY ["QUERY PHASE — every request"]
        D[Skin scores\nfrom Perfect Corp] -->|build query| E[Natural language\nquery string]
        E -->|embed — same model| F[Voyage AI\nvoyage-4-lite]
        F -->|cosine similarity| G[(Supabase\npgvector)]
        G -->|top-k matches| H[Retrieved\nproducts]
    end

    subgraph GENERATE ["GENERATE — grounded only"]
        H -->|inject as context| I[DeepSeek\ndeepseek-chat]
        I --> J[Personalised routine\nno hallucination]
    end

    C -.->|same vector space| G

Loading

The same embedding model is used at both index time and query time. Using different models would place vectors in different semantic spaces — similarity scores would be meaningless.

---

Full system architecture

flowchart TD
    U[User uploads selfie] --> API[FastAPI\nPOST /analyse]
    API --> PC[Perfect Corp\nSkin Analysis API]
    PC --> SC[14 concern scores\n+ skin type]

    SC --> T{Python safety triage\nnot the LLM}
    T -->|score ≥ 0.85| REF[Referral card\nLLM never called]
    T -->|0.40 – 0.85| MOD[Recommend\n+ derm nudge]
    T -->|score < 0.40| MIL[Full OTC\nrecommendation]

    SC --> Q[Build natural\nlanguage query]
    Q --> EMB[Voyage AI embed\n1024-dim]
    EMB --> VEC[(Supabase pgvector\ncosine search)]
    VEC --> RET[Top-k matched\nproducts]

    MOD --> LLM
    MIL --> LLM
    RET --> LLM[DeepSeek\ndeepseek-chat]
    LLM --> RES[Morning + evening\nroutine JSON]

    RES --> FE[Vercel frontend]
    REF --> FE

Loading

---

RAG knowledge base

Detail	Value
Products indexed	100
Embedding model	Voyage AI voyage-4-lite
Vector dimensions	1024
Database	Supabase pgvector
Similarity metric	Cosine
Top-k retrieved	3 per query
Query source	Skin concern scores → natural language

---

Safety design

The triage runs in Python before the LLM is called. Severe concerns are never sent to the LLM.

Tier	Score	Action
Mild	0 – 0.40	Full OTC recommendation
Moderate	0.40 – 0.85	Recommendation + 8-week derm nudge
Severe	0.85+	Referral card only — LLM skipped

LLM system prompt hard limits: no diagnoses, no prescriptions, only reference retrieved products.

---

Stack

Layer	Choice
Backend	FastAPI (Python) — Google Cloud Run
Skin analysis	Perfect Corp HD skin-analysis API
Embeddings	Voyage AI voyage-4-lite (1024-dim)
Vector store	Supabase pgvector
LLM	DeepSeek deepseek-chat
Frontend	HTML — Vercel

---

Local setup

git clone https://github.com/rkchellah/Confidence.git
cd Confidence
pip install -r requirements.txt
cp env.example .env.local
# Fill in all values

# Seed the RAG knowledge base — run once
python scripts/build_product_db.py

# Start the backend
uvicorn backend.main:app --reload

# Open frontend/index.html in your browser

---

Project structure

confidence/
  backend/
    perfect_corp.py      — Perfect Corp API client (async upload → task → poll)
    rag_products.py      — Voyage AI embed + Supabase pgvector retrieve
    routine_generator.py — DeepSeek + 3-tier triage + structured JSON
    main.py              — FastAPI routes + CORS
  frontend/
    index.html           — Landing page + upload/results
  scripts/
    build_product_db.py  — RAG knowledge base seeding (run once)
    sample_products.json — 100 product knowledge base
  Dockerfile
  requirements.txt