week 7 project agentic shopping assistant

week7/community_contributions/norbert-wakanda/project.ipynb

@@ -0,0 +1,30 @@
+{
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "id": "e9466cd3",
+      "metadata": {},
+      "source": [
+        "# Week 7 Project\n",
+        "\n",
+        "The **Jumia AI Product Advisor** is an agentic shopping assistant that leverages a fine-tuned Llama 3.2-3B model, a Retrieval Augmented Generation (RAG) knowledge base, and live web search. It estimates fair product prices and evaluates deals on Jumia's e-commerce platform. The system is orchestrated by Claude Haiku in a tool-calling loop and presented through a Gradio chat UI.\n",
+        "\n",
+        "**Link to project:**  \n",
+        "[Open in Google Colab](https://colab.research.google.com/drive/1d2OqUJPzqZT3LILuWip8xdb0TKEf0roY#scrollTo=aql4NXtCmVWe)"
+      ]
+    }
+  ],
+  "metadata": {
+    "kernelspec": {
+      "display_name": "llms2",
+      "language": "python",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python",
+      "version": "3.13.11"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 5
+}

week7/community_contributions/norbert-wakanda/readme.md

@@ -0,0 +1,75 @@
+# Jumia AI Product Advisor
+
+An agentic AI shopping assistant for Jumia, Africa's leading e-commerce platform. The agent combines a fine-tuned Llama pricing model, a RAG knowledge base, and live web search to help users evaluate product prices and find deals.
+
+## How It Works
+
+The agent (powered by Claude Haiku) orchestrates four tools in a reasoning loop:
+
+1. **`llama_price_estimate`** — Runs a fine-tuned Llama 3.2-3B model (loaded in 4-bit NF4 quantization) to estimate a product's fair market price from its description.
+2. **`rag_search`** — Queries a local ChromaDB vector store built from Jumia markdown documents (products, policies, services). Chunks are generated and reranked by Claude for relevance.
+3. **`web_search`** — Uses DuckDuckGo to find real-time listings and prices on Jumia or other African e-commerce platforms.
+4. **`compare_prices`** — Compares the Llama estimate against a real listing price and returns a deal verdict (great deal → overpriced).
+
+A Gradio UI wraps the agent with a chat interface and a one-click knowledge base initializer.
+
+## Stack
+
+| Component | Technology |
+|---|---|
+| Orchestration LLM | `claude-haiku-4-5` (Anthropic) |
+| Pricing model | Fine-tuned `meta-llama/Llama-3.2-3B` via PEFT/LoRA |
+| Vector store | ChromaDB (persistent) |
+| Embeddings | `all-MiniLM-L6-v2` (HuggingFace) |
+| Web search | DuckDuckGo (`langchain-community`) |
+| UI | Gradio |
+| Runtime | Google Colab (GPU) |
+
+## Setup (Google Colab)
+
+This project is designed to run on **Google Colab** to take advantage of a free CUDA GPU for the Llama pricing model.
+
+**1. Open the notebook in Colab:**
+
+[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1d2OqUJPzqZT3LILuWip8xdb0TKEf0roY#scrollTo=aql4NXtCmVWe)
+
+**2. Download the knowledge base:**
+
+The Jumia knowledge base (markdown files covering products, pricing, company info, employees, and contracts) is hosted on Google Drive:
+
+📁 [Download knowledge-base](https://drive.google.com/drive/folders/1Yq8lk0Utykj8gQ1lkqqWkvkw9GfOEgja?usp=sharing)
+
+Mount your Google Drive in Colab or upload the `knowledge-base/` folder directly to your Colab session so it sits alongside the notebook:
+
+```
+knowledge-base/
+├── company/
+├── contracts/
+├── employees/
+├── pricing/
+└── products/
+```
+
+**3. Set your API keys** in a `.env` file or as Colab secrets:
+
+```
+ANTHROPIC_API_KEY=your_key
+HF_TOKEN=your_huggingface_token
+```
+
+**4. Install dependencies:**
+
+```bash
+pip install torch transformers peft bitsandbytes gradio chromadb \
+            langchain langchain-anthropic langchain-huggingface \
+            langchain-community pydantic tqdm python-dotenv
+```
+
+## Running
+
+Run all cells in the notebook. The Gradio app will launch with a public share link. Click **Initialize / Refresh Knowledge Base** on first run to chunk, embed, and index the knowledge base into ChromaDB.
+
+**Example prompts:**
+- *Is a Samsung Galaxy S24 for $400 a good deal?*
+- *What are Jumia Express delivery fees?*
+- *How does JumiaPay work?*