NLP_project (Spring 2024)

This is a repo for the NLP class project (Retrieval-augmented generation).

The main instruction following: https://github.com/forrestbao/nlp-class/blob/master/projects.md

Team members

Dongyoun Kim, Daeun Kim

Recent updates

• 2024.5.1 upload.py and query.py
• update the upload.py and query.py as composed of simple functions instead of class in each file.
• This is for the gui.py
• 2024.4.15 query
• updated the upload.py file.

# line: 124 - adding {'text': node.text} in upload.py
vectors=[(node.metadata['file_name'][:2]+node.metadata['page_label'], emb, {'text': node.text}) for node, emb in zip(nodes,embedding)],
                    namespace=self.name_space)

Objectives

Build the Retrieval-augmented generation (RAG)

Tech stack

1. Pipeline: [LlamaIndex] (https://www.llamaindex.ai/)
2. VectorDB: [PineCone] (https://www.pinecone.io/)
3. UI: Funix.io

Demo video link

• 1st demo (upload.py): https://youtu.be/m0bHtik0rVs?si=d6gJkUau4J8j9pRF
• 2nd demo (query.py): https://youtu.be/nXHYhM7QiWw?si=S1J_yPOGxLKiBfht
• 3rd demo (gui.py): https://youtu.be/rpEhpPKHfok?si=_C1hN5jv7mMNk0PR

Project Structure

Data repository

.
|-- documents/ # documentation files (Every input have to store it in doucments)
|-- demos/
| README.md
|-- demos/ # videos demos
| upload.py
| query.py
| requirement.txt
| README.md

• documents: This folder contains input files (pdf)
• demos: This folder contains the demo short video
• upload.py: This code reads the file or files in the default folder('documents'), and indexes and stores the file/files in a vector DB ('Pipecone').
• requirement.txt: A list of packages or libraries needed to work on a project that can all be installed with the file.

RAG with gui

Final task

Task: Build a GUI app for all steps above. Include README file and a short video demo to show the usage and completion of your GUI app.

1. Upload pdf: parameters - chunk size and chunk overlap
2. Request the query: parameter - top-k

Implementation

funix gui.py 

Getting Started

KEY setting

#line 47 -  
#Enter the API keys for accessing Pipecone
os.environ["PINECONE_API_KEY"] = 'xxxxxx..' #Enter YOUR KEY
os.environ["PINECONE_ENV"] = "gcp-starter" #Enter YOUR environment in Pipecone(DB)

Sub-tasks1

Task: March 20: PDF upload and indexing via command line done. Include README file and a short video demo to show the usage and completion of your command line tool. e.g., python upload.py --pdf_file=example.pdf to add one PDF file each time.

Steps of the sub-task

1. Read the file : Load and read pdf file + pre-processing for replacing consecutive spaces, newlines and tabs in the file.

2. Chunk: This project chunks the pdf file based on the sentence - chunk size and chunk overlap

3. Embedding: This project use a simple embedding model (bge-small-en-v1.5) from hugging face. Because the OpenAI API (GPT) is subject to rate limitation error.

4. Upload(Upsert) the data

Arguments

    parser = argparse.ArgumentParser(description= 'Process the pdf file for uploading the file to Pinecone (Vector DB)')
    parser.add_argument('--file_name', type=str, default= None, help='A path of input file')
    parser.add_argument('--chunck_size', type=int, default=200, help='Enter the chunck size over 100 range')
    parser.add_argument('--chunck_overlap', type=float, default=0.25, help='The portion of the overlap chunks: 25% = 0.25 range[0,1]')
    parser.parse_args()

1. Single file upload

The file have to store in 'documents' folder.

>>> python upload.py --file_name sample.pdf

2. With optional arguments

>>> python upload.py --name_space test_case --chunck_size 100 --chunk_overlap 0.20 # overlap = 100*0.2 = 20.

• '--file_name': [Optional] Enter the PDF file name. The file have to store in 'documents/'. If do not enter it, the code will Reads the files in the documents folder

  >>> python upload.py --file_name sample.pdf
  

• '--chunk_size': [Optional] Enter the chunk size as integer. The default is 200

  >>> python upload.py --chunk_size 100
  

• '--chunk_overlap': [Optional] Enter the ratio of chunk overlap ranging [0,1]. The default is 0.25

  >>> python upload.py --chunck_overlap 0.25
  

Sub-tasks2

Task: Answer generation based on user queries via command line finish. Include README file and a short video demo to show the usage and completion of your query tool. e.g., python query.py --question="What is the meaning of life?" to get an answer.

Steps of the sub-task

1. Read the query : Read a query + embedding the query with the same model
2. Retrieval k-top chunks: The default of k-top: 5
3. Language Model: a small pre-trained language model by using Llma-CPP with Hugging face- llama-2-13b-chat.Q4_0.gguf

Arguments

    parser.add_argument('--question', type=str, default= None, required=True, help='A query')
    parser.add_argument('--top_k', type=int, default=5, help='top_k')

1. Answer generation with (top-k= 5) based on the query

>>> python query.py --question "What is the attention model?"

2. Answer generation with (user-defined top-k= 5) based on the query

>>> python query.py --question "What is the attention model?" --top_k 8