README

This is the repository for all the material of the Retrieval Augmented Generation training. Here you can find everything you need to deploy a simple RAG application based on the Solr search engine and an OpenAI LLM.

Requirements

To execute the project you need:

• Docker
• Python
• An access to the OpenAI APIs

Repository content

• chunking: contains the chunk.py Python script which takes in input a JSON file of Solr documents and produces a JSON file containing one Solr document for each generated chunk.
• data: contains input data and related Python scripts.
  • create.py: Python script to generate a JSON file of Solr documents.
  • documents_10k.tsv: input text for Solr documents.
  • solr_documents.json: JSON file of Solr documents.
  • solr_documents_with_chunks.json: JSON file of Solr chunked documents.
• docker-solr: contains the Docker file and the solr configuration.
• neural: contains the LLM related classes.
• solr-tools:
  • index_documents.py: Python script to index documents in the Solr intance.
• Solr.py: Solr class to instantiate a Solr instance, index documents and execute queries.
• SolrRetriever.py: SolrRetriever class to generate and retrieve vectors from Solr.
• config.yml: configuration file for the uvicorn application that implements RAG.
• globals.py: global variables.
• main.py: uvicorn application implementing the RAG pipeline.

Configuration

Before proceeding, change the configuration file accordingly to your use-case.

# Solr url where the data is indexed
solr_url: http://localhost:8984/solr/rag_index

# Server port where the langchain server runs
server_port: 8000

# If you set  allenai/scibert_scivocab_uncased it is going to download it at runtime.
# If you have already downloaded the model, you can set here the local path
pretrained_model_path: allenai/scibert_scivocab_uncased

# Min size (number of characters) of a chunk to be indexed 
# All the chunks with fewer characters will be discarded
min_chunk_size: 500

# RAG endpoint
# All the fields are mandatory
endpoint:
  - 
    # OpenAI model name
    model_name: "gpt-3.5-turbo-1106"
    # The name of the endpoint. The server will be run at the address http://URL:<server_port>/<endpoint_name>/playground
    endpoint_name: "gpt_3_5_rag"
    # Number of rows to retrieve for the knn phase
    knn_rows: 10 
    # Number of rows to retrieve for the bm25 phase
    bm25_rows: 10

Installation

To run Solr with Docker, follow the following instructions:

cd docker-solr;
docker-compose up;

Solr will be available at https://localhost:8984/

To run Python scripts, install the requirements:

pip install -r requirements.txt;

Generate documents

You can skip this step if you want to use the already provided material. To generate documents:

cd data;
python create.py

To generate chunked documents:

cd chunking;
python chunk.py

Index documents

cd solr-tools;
python index_documents.py "<path to>/RAG-training/data/solr_documents_with_chunks.json"

Run RAG server

export OPENAI_API_KEY="<your-api-key-here>"
python main.py

Usage

Start Solr:

cd docker-solr;
docker-compose up;

Start the RAG server:

python main.py

Make queries at: http://0.0.0.0:8000/gpt_3_5_rag/playground/