Food RAG - Production Hybrid RAG System

A production-grade food recommendation API built with hybrid BM25 + vector search, metadata filtering, cross-encoder reranking, semantic caching, and RAGAS-style evaluation. The system uses a local Ollama model for grounded answer generation and LLM-as-judge generation evaluation.

---

Demo:

Demo Link

---

Project Structure

food-rag/
|-- api/
|   |-- app.py          # FastAPI app + lifespan
|   |-- routes.py       # /health, /recommend, /eval
|   `-- schemas.py      # Pydantic request/response models
|-- data/
|   |-- FoodDataSet.json
|   `-- eval_data/
|       `-- test_samples_50.json
|-- scripts/
|   `-- preprocess_usda_fooddata.py
|-- src/
|   |-- cache.py        # Exact + semantic cache
|   |-- config.py       # Settings from .env
|   |-- data_loader.py  # Dataset loading + normalization
|   |-- evaluator.py    # Retrieval + generation eval metrics
|   |-- filters.py      # Nutrition/query filter parser
|   |-- hybrid_retriever.py  # BM25 + ChromaDB + RRF
|   |-- query_transformer.py # Router, multi-query, step-back
|   |-- rag_pipeline.py # Pipeline orchestrator
|   |-- reranker.py     # Cross-encoder + MMR
|   `-- vector_store.py # ChromaDB interface
|-- main.py
|-- requirements.txt
`-- .env.example

---

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
ollama pull llama3.2:3b

Make sure Ollama is running before starting the API:

ollama serve

On Windows PowerShell, activate the environment with:

.\.venv\Scripts\Activate.ps1

---

Dataset

The app expects the production dataset here by default:

data/FoodDataSet.json

Default .env setting:

DATASET_PATH=./data/FoodDataSet.json
OLLAMA_BASE_URL=http://127.0.0.1:11434
OLLAMA_MODEL=llama3.2:3b

Each food record should support RAG fields such as:

{
  "id": "2727567",
  "food_name": "Chicken thigh",
  "aliases": ["chicken thigh"],
  "semantic_text": "Food Name: Chicken thigh...",
  "metadata": {
    "calories": 221,
    "protein_g": 18.6,
    "low_sodium": true,
    "high_protein": true
  },
  "tags": ["high-protein", "low-sodium"]
}

---

USDA Preprocessing

If you need to regenerate the dataset from a USDA-style JSON file:

python scripts/preprocess_usda_fooddata.py --input "C:/path/to/FoodData_Central_foundation_food_json_2026-04-30.json"

Output:

data/FoodDataSet.json

The preprocessor normalizes food names, extracts nutrition fields, computes missing calories, generates nutrition tags, creates aliases, builds semantic text, and removes duplicates.

---

Running The API

python main.py --serve

Or run directly with reload:

uvicorn api.app:app --host 0.0.0.0 --port 8000 --reload

Interactive docs:

http://127.0.0.1:8000/docs

---

API Endpoints

GET /health

Returns pipeline readiness and cache stats.

POST /recommend

{
  "query": "healthy food under 300 calories with low sodium",
  "n_results": 5,
  "use_cache": true
}

Example query ideas:

{
  "query": "high protein low fat foods",
  "n_results": 3,
  "use_cache": true
}

{
  "query": "foods for muscle recovery",
  "n_results": 3,
  "use_cache": true
}

DELETE /recommend/cache

Clears recommendation cache, semantic cache, and route cache.

POST /eval

[
  {
    "query": "show me protein rich food for a post workout meal",
    "ground_truth_answer": "Chicken thigh, pollock, and pine nuts are relevant high-protein options for post-workout recovery.",
    "relevant_doc_ids": ["2727567", "333476", "2346392"]
  }
]

---

CLI Usage

# Single query
python main.py "healthy food under 300 calories with low sodium"

# Interactive mode
python main.py

# Rebuild ChromaDB index
python main.py --rebuild-index "high protein low fat foods"

# Disable cross-encoder for faster testing
python main.py --no-cross-encoder "foods rich in potassium"

# Run evaluation
python main.py --eval data/eval_data/test_samples_50.json

# Cache stats
python main.py --cache-stats "low sodium healthy snacks"

---

Query Routing

The router chooses one strategy per query:

Strategy	Used for
NORMAL_RETRIEVAL	Exact food names, direct ingredient searches, simple nutrition lookups
MULTI_QUERY	Nutrition constraints, calorie filtering, diet searches, ingredient combinations
MULTI_QUERY_STEPBACK	Goal-based or recommendation-style queries like muscle recovery or digestion
CLARIFICATION	Queries too vague to retrieve reliably
REJECTION	Non-food topics

Examples:

apple pie calories -> NORMAL_RETRIEVAL
healthy foods under 300 calories -> MULTI_QUERY
foods for muscle recovery -> MULTI_QUERY_STEPBACK
protein -> CLARIFICATION
weather tomorrow -> REJECTION

---

Evaluation

Retrieval metrics:

• Recall@K
• Precision@K
• MRR
• Hit Rate

Generation metrics use LLM-as-judge:

• Faithfulness
• Answer relevancy
• Context precision
• Context recall
• Answer correctness

Run:

python main.py --eval data/eval_data/test_samples_50.json

---

Architecture

POST /recommend
      |
      v
Query Router
      |
      +--> CLARIFICATION / REJECTION
      |
      +--> NORMAL_RETRIEVAL
      +--> MULTI_QUERY
      +--> MULTI_QUERY_STEPBACK
      |
      v
Hybrid Retrieval
  - Dense: ChromaDB vector search
  - Sparse: BM25 keyword search
  - RRF fusion
      |
      v
Cross-encoder Reranker + MMR Diversity
      |
      v
Grounded Generation with Ollama llama3.2:3b
      |
      v
Recommendation Cache + Semantic Cache + Route Cache

---

Metadata Filtering

The dataset metadata is optimized for filtering:

where = {
    "$and": [
        {"calories": {"$lte": 300}},
        {"high_protein": {"$eq": True}},
        {"low_sodium": {"$eq": True}},
    ]
}

Supported nutrition-aware searches include:

• healthy food under 300 calories
• high protein low fat foods
• vegan foods rich in iron
• foods rich in potassium
• low sodium healthy snacks
• diabetic-friendly foods
• foods for muscle recovery