.png){kind=link}
.png){kind=link}
⚙️ Retrieval-Augmented Generation (RAG) over the Indian Food Composition Tables (IFCT 2017)
🧠 Combining NLP, Information Retrieval, and Nutrition Science for culturally aware dietary guidance.
.png)
While several global calorie-tracking tools exist, most fail to adapt to the diverse Indian cuisine, which involves complex preparation styles, ingredients, and regional variations.
Bite2Burn addresses this gap by integrating AI, Natural Language Processing (NLP), and Knowledge Retrieval Systems to enable accurate calorie estimation and meal planning based on the Indian Food Composition Tables (IFCT) developed by the National Institute of Nutrition (ICMR).
This project blends machine learning, data engineering, and human-computer interaction to deliver a seamless, interactive chatbot experience — empowering users to make informed dietary decisions and achieve their health goals effectively.
🧩 This project lies at the intersection of:
- 🍱 Nutritional Informatics — structuring and normalizing food composition knowledge
- 🔍 Information Retrieval — dense vector similarity + multi-granularity indexing
- 🧾 Natural Language Generation — LLM-driven summarization and answer synthesis
- 🛠️ Applied Data Engineering — PDF parsing, OCR, table normalization
💡 Contribution: Toward precision nutrition, dietary decision support, and culturally aligned health AI.
Published by: National Institute of Nutrition (ICMR)
Scope: 528 food items · 20 food groups · 150+ nutritional parameters
Granularity: Region-wise variations (North, South, East, West, Central, NE India)
| 🍚 Attribute | 📊 Description |
|---|---|
| Total Foods | 528 |
| Food Groups | 20 |
| Nutrients Analyzed | >150 |
| Units | kJ per 100g edible portion |
| Data Source | National Institute of Nutrition (ICMR) |
| Uniqueness | Lab-analyzed values with cultural food prep details |
The pipeline combines layout-aware document processing, semantic retrieval, and generative reasoning for food-level understanding.
| 🔹 Stage | 🧰 Technique / Tool | 🎯 Purpose |
|---|---|---|
| 🧾 PDF Partitioning | Unstructured + layout inference | Structured extraction of tables & titles |
| 🔤 OCR (conditional) | Tesseract |
Convert embedded raster text |
| 🗂️ Categorization | Type inspection (Table vs CompositeElement) |
Downstream semantic handling |
| 🧠 Summarization | Mistral-7B-Instruct (instruction-tuned) |
Condense verbose tables for efficient embedding |
| 🔡 Embeddings | FastEmbed |
Dense vector semantic search (low-latency) |
| 🔍 Multi-Vector Retrieval | LangChain MultiVectorRetriever |
Link summary → raw table/text for rich context |
| 🧱 Vector Store | ChromaDB |
Persistent similarity index |
| 💬 Answer Generation | Mistral / (planned) Llama 3.1 |
Natural language synthesis grounded in retrieved context |
| ⚖️ Personalization Logic | Derived caloric equations (future extension) | Tailored intake/expenditure guidance |
| User Query | Bite2Burn AI Response (Example) |
|---|---|
| “What’s the calorie content of cooked rice?” | Cooked rice contains approximately 130 kcal per 100g according to IFCT (2017). The energy values vary slightly between East and South India. |
| “How many steps do I need to burn 500 kcal?” | For a 70kg individual walking at 5 km/h, approximately 8,500–9,000 steps are needed to burn 500 kcal. |
.png)
| 🧩 Layer | ⚙️ Technology / Library | 🎯 Purpose |
|---|---|---|
| 🧠 LLMs | Mistral 7B, Llama 3.1 |
Summarization & Question Answering |
| 🗃️ Vector DB | ChromaDB |
Semantic similarity search & retrieval |
| 🔡 Embedding | FastEmbed |
Text-to-vector encoding |
| 📄 PDF Processing | Unstructured, YOLO |
Text & table extraction |
| 🎨 Frontend | React, Tailwind CSS |
Interactive UI/UX |
| ⚙️ Backend | Flask |
RESTful API services |
| 📊 Visualization | Recharts |
Nutrient charts & data plots |
| 🌐 Communication | Axios |
HTTP request handling between client and server |
💡 Outputs are informational and not medical advice.