Travel Destination Recommender

This is an LLM-based recommender for travel destinations.

Set-up

1. Run the following command to set up your OpenAI API key:

   cp config_template.py config.py
   

   Open config.py and replace "YOUR_API_API_KEY" with your actual API key.

2. Run the following command to set up enviroment:

   conda env create -f environment.yml
   

Embedder

The embedder encodes destination text data scraped from Wiki-travel into high-dimensional vector representations, using either a GPT-based model or a Sentence-Transformer based model.

Usage

Run the following command:

python -m src.Embedder.embedderRunner -d data/clean_destination_air_canada_xml -o output --split_type <sentence_or_section> --emb_type <gpt_or_st>

• -d, --data_path: Path to the directory containing text files.
• -o, --output_type: Path where embeddings should be saved.
• --split_type: The type of text splitting to apply before embedding. Available types: 'sentence' or 'section'.
• --emb_type: Specify the type of the embedder. Available types: 'gpt' or 'st'.

Query Processor

The query processor analyzes a given user query by breaking it down into a list of aspects and then applies different processing strategies to each aspect to better capture and reflect the user's intent:

• Reformulation: Rewrites the aspect description.
• Expansion: Generates three improved descriptions and concatenates these with the original aspect.
• Elaboration: Generate a detailed paragraph that elaborates on the original aspect.

Usage

Run the following command line:

python -m src.QueryProcessor.queryProcessorRunner -i data/queries.txt -o output --mode <reformulate_or_expand_or_elaborate>

• -i, --input_path: Specifies the path to the input file containing the user queries. This file should be in text format, with each query on a new line.

• -o, --output_dir: Defines the directory where the processed queries will be saved.

• --mode: Determines the processing strategy to apply to each query. There are three available options: 'reformulate', 'expand', or 'elaborate'.

Retriever

The retriever extracts relevant text chunks from destination texts based on processed user queries.

Usage

Run the following command line:

python -m src.Retriever.retrieverRunner -q query/path -e embeddings/path -o output/path --emb_type <gpt_or_st>

• -q, --query_path: Path to the input file containing processed queries. This should be a pickle file that holds the processed query from QueryProcessor module.
• -e, --embedding_dir: Directory containing the embeddings that represent the textual data of the destination texts.
• -o, --output_path: Path where the retrieval results will be stored. The output should be a JSON file that provides scores and relevant text chunks for each city and query aspect.
• --emb_type: Specify the type of the embedder. Available types: 'gpt' or 'st'.