♻️ LLM-Based Sustainability Assistant for EPD Analysis

This project is an AI-powered system for analyzing and comparing Environmental Product Declarations (EPDs). It integrates structured data, semantic embeddings, and large language models (LLMs) to extract, enrich, and query sustainability-related information.

---

📌 Project Goals

• Ingest and store EPD data from XML and PDF sources
• Enrich incomplete product information using LLMs
• Enable semantic search and comparison across indicators, materials, and categories
• Provide a fast, interactive frontend for analysis and Q&A

---

✅ MVP (Minimum Viable Product)

📥 Data Collection & Storage

• Retrieved EPD data using Selenium, parsed XML.
• Stored structured product and indicator data in PostgreSQL.

🧠 Local LLM Experiments

• Tested Mistral-7B-Instruct (quantized) locally.
• Built a basic React + FastAPI full-stack prototype.
• Used all-MiniLM-L6-v2 for initial embeddings.
• Switched to Ollama with nous-hermes2 due to poor results and slow inference (60s per request).

---

🧪 Development Phase

⚡ Inference Optimization

• Deployed a prebuilt Ollama container on RunPod using:
  • RTX 2000 Ada GPU, 16 GB VRAM, 31 GB RAM, 100 GB disk
• Benchmarked multiple LLMs:
  • Mistral (7B), LLaMA3 (8B), Gemma, Qwen, Phi-3
• Selected Mistral for consistent performance.

🧱 Text Processing & Embedding

• Extracted text manually from PDFs and saved as .txt.
• Programmatically chunked into 3–4 sentence segments.
• Checked chunks for semantic relevance using LLM.
• Embedded with bge-small-en-v1.5 (384 dimensions).
• Stored in PostgreSQL using pgvector.

🧽 Data Enrichment & Cleanup

• Filled missing English descriptions using LLM-based German translation.
• Inferred material and use-case from unstructured text.
• Generated statistics (mean, min, max, range) for indicators by category for fast retrieval.

---

🚀 Final System Architecture

💻 Frontend – React

• Filters products by:
  • Indicator, Category, Material, Use Case
• Retrieves:
  • Structured product/indicator data from PostgreSQL
  • Embedded chunks from EPD XML and theory PDFs (via pgvector)
• Sends questions to FastAPI for:
  • LLM-based classification
  • Routed LLM answers

🧠 Backend – FastAPI

• Interfaces with both the database and LLM.
• Handles:
  • PostgreSQL queries for structured + embedded data
  • pgvector semantic retrieval
  • LLM prompt generation and routing

🗄️ Database – PostgreSQL + pgvector

• Stores:
  • Parsed XML product data
  • Chunked and embedded EPD text and theory documents
• Enables:
  • Standard SQL queries
  • Vector similarity search for semantic context

🤖 LLM Backend – RunPod/Ollama

• Hosts Mistral (7B quantized)
• Handles:
  • Question classification
  • Answer generation using retrieved context
• Communicates with FastAPI via prompt/response API

---

📎 Tech Stack

Component	Tech
Frontend	React
Backend	FastAPI
Database	PostgreSQL + pgvector
Embeddings	bge-small-en-v1.5
LLMs	Mistral, LLaMA3, etc. via Ollama on RunPod

---

📂 Folder Structure (Suggested)