This project was developed for the Hackapizza Kaggle Competition by Stefano Iannicelli and Ettore Caputo.
The goal? To create a super-smart solution for context-aware question answering using structured menu data .
We focused on efficient token usage, precise data extraction, and multi-phase query processing .
-
📂 Structured Data Processing:
- Splits menu files into structured chunks like headers and dishes
- Tags dish names with
<dish></dish>for easy spotting - Rebuilds tables and decodes Roman numerals using regex
-
🧠 Multi-Expert System:
- Tech Expert: Understands fancy cooking methods from the Galactic Code
- Distance Expert: Finds restaurants by planetary distances
- Menu Header Expert: Filters based on restaurant metadata
- Menu Corpus Expert: Dives deep into the menu content for dish details
-
🔎 Boolean Query Processing:
- Transforms user queries into boolean expressions
- Filters menu data with structured keyword logic
- Ensures super precise answers every time
-
⚙️ Token Efficiency:
- Minimizes dependence on LLMs thanks to boolean smarts
- Makes every token count for context-aware replies
- 🔑 Keyword Extraction – Pulls out the important bits from the question
- 🛠️ Query Reformulation – Turns them into boolean expressions
- 🧠 Expert Activation – Different experts handle their part of the query
- 📚 Boolean Search – Finds the matching data
- 🍽️ Final Answer Extraction – Grabs dish names straight from the filtered content
| 🧪 Configuration | 🎯 Score (%) |
|---|---|
| Menu Expert Only | 63.5 |
| + Distance Expert | 66.7 |
| + Tech Expert | 76.5 |
- 📐 Rigid Boolean Model – Very structured queries; even small keyword slips can cause issues
- 🧠 Tech Expert Optimization – Currently sends the whole Galactic Code to the LLM. Switching to chunk-based retrieval could save tons of tokens!
📄 Want more details? Check out the project PDF! Thanks for reading! 🙌