🥦 Precision Nutrition AI — Assignment 2

Qatar University · CMPE 682/783 · Intelligent Systems
Upgrade from A1 (pure prompting) → A2 (real RAG + real ReAct agent)

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│                     Streamlit UI (app.py)                   │
│   Chat │ Evaluate │ Profile │ About                         │
└───────────────────────┬─────────────────────────────────────┘
                        │
              ┌─────────▼──────────┐
              │  safety.py         │  ← Pre-screening gate
              │  (runs FIRST)      │
              └─────────┬──────────┘
                        │
              ┌─────────▼──────────┐
              │  agent.py          │  ← ReAct loop (qwen3.5:4b)
              │  Semantic cache    │    max 5 steps
              └──┬──────────┬──────┘
                 │          │
    ┌────────────▼─┐   ┌────▼────────────┐
    │  rag_engine  │   │   tools.py      │
    │  ChromaDB    │   │   7 real tools  │
    │  MiniLM-L6   │   └─────────────────┘
    └──────┬───────┘
           │
    ┌──────▼───────────────┐
    │  nutrition_knowledge │
    │  10 clinical docs    │
    │  (USDA/ADA/AHA/NIH…) │
    └──────────────────────┘

---

Setup

1 — Python dependencies

pip install -r requirements.txt

2 — Ollama + model

# Install Ollama: https://ollama.com/
ollama pull qwen3.5:4b
ollama serve          # keep running in a separate terminal

3 — Run

cd Local_Precision_Nutrition_AI/
streamlit run app_streamlit.py

The ChromaDB vector store is built automatically on first launch.

---

Features

Feature	A1 Baseline	A2 Enhanced
RAG	❌	✅ Real ChromaDB + sentence-transformers
Tools	❌	7 real tools
Food DB	❌	10+ items
Safety	Reactive (in prompt)	+ Pre-screening guardrail module
Citations	❌	✅ Source labels in responses
Multi-turn	❌	✅ + user profile injection
Evaluation	by non-expert humans	25-case LLM-as-judge suite (7 criteria)
UI	❌	4 tabs (Chat, Evaluate, Profile, About)
Caching	❌	✅ Semantic cache
Tool timeout	❌	✅ 30 s via ThreadPoolExecutor
User abort	❌	✅ Stop button

Extra Features (beyond assignment)

• Safety guardrails — eating disorder detection → mandatory referral; CKD/pregnancy/paediatric flags
• User profile — auto-injected into every prompt (age, weight, goal, conditions)
• LLM-as-judge evaluation — 5 criteria, 25 test cases, A1 baseline comparison
• 4 Plotly charts — before/after bar, category breakdown, radar, tool frequency
• 4-tab Streamlit UI — Chat / Evaluate / Profile / About

---

Tools (7)

Tool	Description
calculate_tdee_bmi	Mifflin-St Jeor BMR → TDEE → BMI category + deficit/surplus
calculate_macro_targets	Macro split by goal (weight_loss / maintenance / muscle_gain / athletic)
get_food_macros	35+ USDA-calibrated foods, fuzzy match, per-serving scaling
retrieve_rag_context	Semantic search over 10 clinical documents
check_supplement_safety	NIH UL database (15 supplements, risk assessment)
analyze_meal_nutrition	Parse meal text → sum macros from food DB
calculate_hydration_needs	WHO-based water target + activity + climate adjustments

---

Evaluation

Run from the Evaluate tab in the app, or:

# Standalone
from evaluation import run_evaluation_suite
results = run_evaluation_suite(n_cases=25)

Metrics: accuracy, actionability, safety, groundedness (new), task_completion (new) , citation accuracy (new), retrieval relevance (new) LLM judge: qwen3.5:4b scores each response 1–5 per criterion.

---

Project Structure

nutrition_ai/
├── app_streamlit.py       # Streamlit UI (4 tabs)
├── agent.py               # ReAct loop
├── tools.py               # 7 tools
├── rag_engine.py          # ChromaDB RAG
├── nutrition_knowledge.py # 10 clinical docs
├── safety.py              # Pre-screening guardrails
├── evaluation.py          # 25-case eval framework
├── requirements.txt
├── README.md
├── chat_history.json      # auto-created when starting a chat
├── eval_results.json      # auto-created when evaluation finishes
├── user_profile.json      # auto-created on first run
├── Sources/               # Sources directory (for RAG)
└── chroma_db/             # auto-created on first run

---

Troubleshooting

ModuleNotFoundError: chromadb
→ pip install chromadb

ollama.ResponseError: model not found
→ ollama pull qwen3.5:4b

Ollama connection refused
→ Make sure ollama serve is running in another terminal

sentence_transformers slow first load
→ Normal — model downloads (~80 MB) on first run, then cached

---

A1 → A2 Improvements Summary

Criterion	A1 Score	A2 Target
Nutritional Accuracy	4.67 / 5.00	4.88 / 5.00
Practical Actionability	4.07 / 5.00	4.64 / 5.00
Safety & Boundaries	5.00 / 5.00	5.00 / 5.00
Groundedness	—	3.28 / 5.00
Task Completion	4.53 / 5.00	4.96 / 5.00
Citation Accuracy	—	2.84 / 5.00
Retrieval Relevance	—	3.64 / 5.00