Classifies recipe ingredients as vegan and/or keto. Handles the messy real-world stuff: compound ingredients, hidden animal products (L-cysteine, gelatin, carmine), and regional synonyms. Runs locally. No cloud API required.
97.8% word-level parsing accuracy on 8,789 ingredients. ~0.6s per uncached ingredient, 135x faster than the 82s brute-force baseline. >90% cache hit rate in practice, so most ingredients resolve in <10ms.
Four stages in sequence:
Parse:
ingredient-parser-nlpextracts the core ingredient name from noisy recipe text (97.8% word-level accuracy)."3 pounds pork shoulder, trimmed and cut into 2-inch chunks"becomes"pork shoulder".Retrieve: Arctic Text2SQL generates a semantic
LIKEquery against two databases. Prioritizes raw forms over processed variants to avoid contaminated nutritional data. Falls back to RapidFuzz if the semantic query returns nothing.Dual databases: Nutrition facts DB (8,789 ingredients) for keto analysis. Vegan ontology (236+ terms) for detecting hidden animal products that nutrition data misses.
Judge: Qwen 0.5B weighs the retrieved evidence and outputs
is_vegan,is_keto, and a plain-language reason.
135x faster than the previous brute-force API baseline (82s -> 0.6s per ingredient).
graph TD
subgraph "Input"
A["Raw Recipe Ingredient List<br/><i>e.g., '2 tbsp extra virgin olive oil'</i>"]
end
subgraph "1. Parsing Stage"
B["ingredient-parser-nlp<br/>(97.8% Accuracy)"]
A --> B
end
subgraph "2. Semantic Retrieval Stage"
C["Cleaned Ingredient<br/><i>e.g., 'olive oil'</i>"]
D["Arctic Text2SQL Engine"]
E["<b>Nutrition Database</b><br/>(8,789 ingredients)"]
F["<b>Vegan Ontology</b><br/>(236+ terms)"]
G["RapidFuzz Fallback"]
B --> C --> D
D -- "Generates Semantic<br/>LIKE Query" --> E
D -- "Generates Semantic<br/>LIKE Query" --> F
C -.->|If Semantic Search Fails| G
G --> E
G --> F
end
subgraph "3. Reasoning & Caching Stage"
H["Retrieved Evidence<br/>- Nutritional Data<br/>- Animal Product Flags"]
I["Qwen 0.5B Judge Model"]
J["Redis Cache"]
K["Final Classification<br/><i>{is_keto: true, is_vegan: true}</i>"]
E --> H
F --> H
H --> I
I -- "Writes to & Reads from" --> J
I --> K
end
style A fill:#FFF,stroke:#333,stroke-width:2px
style B fill:#E3F2FD,stroke:#333,stroke-width:2px
style D fill:#E3F2FD,stroke:#333,stroke-width:2px
style I fill:#E3F2FD,stroke:#333,stroke-width:2px
style E fill:#FFF3E0,stroke:#333,stroke-width:1px
style F fill:#FFF3E0,stroke:#333,stroke-width:1px
style J fill:#D1C4E9,stroke:#333,stroke-width:1px
style K fill:#C8E6C9,stroke:#333,stroke-width:2px
importasynciofromingredient_senseimportClassifierclassifier=Classifier()
recipe_ingredients= [
"3 pounds pork shoulder, cut into chunks",
"2 tbsp extra virgin olive oil",
"1 large onion, chopped",
"4 cloves garlic, minced",
"1 cup almond flour"
]
asyncdefmain():
result=awaitclassifier.classify_recipe(recipe_ingredients)
print(f"Is Vegan: {result['is_vegan']}")
print(f"Is Keto: {result['is_keto']}")
print(f"Reasoning: {result['reasoning']}")
foriteminresult['details']:
print(f"{item['original']} -> {item['parsed']}: vegan={item['is_vegan']}, keto={item['is_keto']}")
asyncio.run(main())Output:
Is Vegan: False
Is Keto: True
Reasoning: Not vegan -- contains pork shoulder. Keto-friendly -- all ingredients are low-carb.
| Metric | Result |
|---|---|
| End-to-end latency (uncached) | ~0.6s / ingredient |
| End-to-end latency (cached) | <10ms / ingredient |
| Cache hit rate | >90% |
| Semantic match quality | 85%+ excellent/good |
| Data loading time | ~1.4s (Polars, 8,789 records) |
Polars, Redis, Ollama (Arctic Text2SQL + Qwen 0.5B), ingredient-parser-nlp, RapidFuzz, MLflow, Numba, uv.
Prerequisites: Python 3.11+, Ollama, Redis running locally.
# Pull models
ollama pull snowflake/arctic-text2sql-artefact
ollama pull qwen:0.5b
# Install
git clone https://github.com/ShovalBenjer/argmax_solution.git
cd argmax_solution
python -m venv .venv &&source .venv/bin/activate
pip install uv
uv pip sync requirements.txt
# Run
python main.pyMIT.