Skip to content

Latest commit

 

History

History
136 lines (88 loc) · 5.78 KB

README.md

File metadata and controls

136 lines (88 loc) · 5.78 KB
Raw

Info

This project is for the 2-day Hackathon organized jointly by Ecole polytechnique, McKinsey and Mistral AI.
Participants are CHEN Yiming, MONTAGNE Thibaud, ZOU Siyuan, ZOU Yuran (by name order).

RetailX-Hackathon

A MVP (Minimum Viable Product) Prototype of RetailX, an intelligent AI assistant powered by Mistral AI designed to provide an online shopping experience comparable to in-store retail.

Key ideas :

  • User-friendly, natural, visible shopping experience;
  • Customized guidance, following the customer's previous choices;
  • Scalable and comprehensible structure.

Demo picture

2

Usage

  1. Create a conda or python venv
conda create -n retailx python=3.10.15

conda activate retailx
  1. Install the dependencies from requirements.txt
pip install -r requirements.txt

  1. Download the dataset Amazon-Sales-Dataset-EDA-94 and copy it to the root folder of the project.

  2. Create a file named "env.py" and add the following items:

file_path = 'Data/amazon.csv'  # by default
api_key = "your_mistral_AI_API_key"
history_file_path = 'history.txt' # by default

pricer_path = 'Personalities/personality_pricer.txt' 
discounter_path = 'Personalities/personality_discounter.txt'
rater_path = 'Personalities/personality_rater.txt'
salesperson_path = 'Personalities/salesperson.txt'
keyword_extractor_path = 'Personalities/keyword.txt'
conselor_path='Personalities/conselor.txt'

ATTENTION: Tested Python version: 3.10.15

  1. Execute in the system terminal gradio app.py, then a link directing to the agent appears in the terminal.

The default link is http://127.0.0.1:7860

Note: Initial Data Preparation may take 30s to 1min.

File Description

Frontend

  • app.py: The main application file that runs the Gradio interface.
  • style.css: Contains custom CSS styles for the Gradio interface.
  • api_frontend.py: The Interface to interact with the backend.

Backend

  • api_backend.py: API for the frontend provided by the backend, also the local test code for backend.
  • functions.py: Functions implementing LLM calls.
  • embed.py: Functions calculating the objects' embeddings
  • api_backend.py: API for the frontend provided by the backend, also the local test code for backend. Contains the main workflow of Backend.
  • functions.py: Auxialiary Functions, implementing embedding comparisons and LLM calls.

Data

  • env.py: Contains environment variables such as file paths and API keys.
  • /Data/amazon.csv: The dataset file containing Amazon sales data.

Miscellaneous

  • requirements.txt: Lists all the dependencies required to run the project.
  • README.md: This file, providing an overview and instructions for the project.
  • mock_request.py: mock data for the frontend.
  • history.txt: A file to store the history of processed data or interactions. Used when testing backend.
  • Personalities: containing System Prompts of all LLM's used in the project.

Technical Details

Front-End

The frontend is powered by the framework gradio. The backend is a simplified synchronous implementation of the system. Instead of leveraging asynchronous programming for handling concurrent requests or external API calls, it processes all incoming requests in a blocking manner.

Back-End

The backend workflow chart looks like this:

2

We use multiple Mistral AI LLM personalities:
The historian extracts a one-phrase description of the user's demand from chat history;
The price/discount/rating filterer extracts any "hard" criteria on those numerical columns, if applicable;
The salesperson generates (based on previous conversations and relevant objects) a pertinent follow-up question to help narrow down.
The suggestionner suggests (based on most relevant items) some items in another category as possible cross-sales.

The actual object-matching is done by local transformer models, by comparing text embeddings.

Cross-sales is done by randomly picking marchandizes from the categories that appear among the most-relevant articles.
We have hard-filter identification mechanisms on numerical features(price,discount rate,rating) and this is easily scalable if there is more similar features.

Model Evaluation

Good points

Mistral AI LLMs generally perform quite well on the given tasks;
Rare hallucinations are observed, even with Large, but detectable and fixable by asking LLM to re-run.

Our chatbot is capable of treating common merchandies search requests, and can raise pertinent questions to narrow down.
The suggestion engine successfully proposes the most relevant items, in a user-friendly way.

Not-yet-good points

Unconventional (not following guidance) user inputs are not always well treated;
Detailed comparison between different encoders or LLM models are not carried out;
There should be a more flexible way to treat 'hard' user-demand-filters, in case there are more numerical columns;
Cross-sales sometimes gives same-category objects because there are not any closely-related category in the database. Scaling up the database may help a lot, but this is not experimented. If we want in the same time top items AND cross-sales, the reaction time is too long. (Too many LLM requests to do).

Future possibilities

A (not implemented but) potentially promising idea is to train a LLM coder to write and auto-execute code snippets that can operate autonomously on the database;
This can allow for more versatile question-answering capabilities related to the database.

A Graph Neural Network that defines item similarity might be able to give more logical cross-sales suggestions.