rag

RAG with Data Prep Kit

This folder has examples of RAG applications with data prep kit (DPK).

Step-0: Getting Started

## Clone this repo
git  clone   https://github.com/IBM/data-prep-kit

cd data-prep-kit
## All commands from here on out, assume you are in project root directory

Step-1: Setup Python Environment

setup-python-dev-env.md

Make sure Jupyter is running after this step. We will use this Jupyter instance to run the notebooks in next steps.

RAG Workflow

Here is the overall work flow. For details see RAG-explained

Step-2: Process Input Documents (RAG stage 1, 2 & 3)

This code uses DPK to

• Extract text from PDFs (RAG stage-1)
• Performs de-dupes (RAG stage-1)
• split the documents into chunks (RAG stage-2)
• vectorize the chunks (RAG stage-3)

Here is the code:

• Python version: TODO
• Ray version: rag_1A_dpk_process_ray.ipynb

Step-3: Load data into vector database (RAG stage 4)

Our vector database is Milvus

Run the code: rag_1B_load_data_into_milvus.ipynb

Be sure to shutdown the notebook before proceeding to the next step

Step-4: Perform vector search (RAG stage 5 & 6)

Let's do a few searches on our data.

Code: rag_1C_vector_search.ipynb

Be sure to shutdown the notebook before proceeding to the next step

Step-5: Query the documents using LLM (RAG steps 5, 6, 7, 8 & 9)

We will use Llama as our LLM running on Replicate service.

5.1 - Create an .env file

To use replicate service, we will need a replicate API token.

You can get one from Replicate

Create an .env file (notice the dot in the file name in this directory with content like this

REPLICATE_API_TOKEN=your REPLICATE token goes here

5.2 - Run the query code

Code: rag_1D_query_llama_replicate.ipynb

Step 6: Illama Index

For comparision, we can use Llama-index framework to process PDFs and query

Step 6.1 - Process documents and save the index into vector DB

Code: rag_2A_llamaindex_process.ipynb

Be sure to shutdown the notebook before proceeding to the next step

Step 6.2 - Query documents with LLM

code: rag_2B_llamaindex_query.ipynb

Tips: Close the notebook kernels, to release the db.lock

When using embedded database, the program maintains a lock on the database file. If the lock is active, other notebooks may not be able to access the same database.

Here is how to close kernels:

• In vscode: Restart the kernel. This will end the current kernel session and release the db.lock
• In Jupyter : Go to File --> Close and shutdown notebook. This will end the current kernel session and release the db.lock