RAG-Pipeline-1

Created the first RAG pipeline

Vertex AI Project

Steps Followed :

1. Create a new Google account.

2. Credits: Received $300 in credits.

3. Enable Vertex AI APIs.

4. Create a bucket in Google Cloud Storage (GCS).

5. Use Vertex AI Workbench:

   • Alternatively, you can use a DataProc cluster with Jupyter Lab as an installed application.
   • Create a Jupyter Notebook using a machine with specific configurations: Vertex AI Workbench

   Note: It takes about 7-8 minutes to bring the Jupyter Notebook up.

Notebooks Created

1. embeddings_gen_and_vector_index_deployment.ipynb

• This notebook generates embeddings for the statements in the uploaded PDFs.
• The embeddings file is uploaded to GCS at the specified file path.
• A vector search index is created using the URI of the embeddings file created in the previous step.
• A Matching Index Endpoint is created and deployed. Deployment takes around 20-25 minutes for a small PDF, but this time duration can vary based on the instance used and the size of the data.

Important: Make sure to pass the following parameters in the deploy_index function, or the deployment will not succeed:

• machine_type=machine_type
• min_replica_count=min_replica_count
• max_replica_count=max_replica_count

2. perform_semantic_search_and_get_output.ipynb

• Initialize the vector search index.
• Generate embeddings for user input.
• Find nearest neighbors for the user input embeddings from the Matching Index Endpoint.
• For example, if you get the IDs of 10 nearest neighbors:
  • Look up these 10 UUIDs in the sentences.json file to retrieve the corresponding sentences for context creation.
• The retrieved 10 sentences are now referred to as the context.
• Create a prompt that injects the above-created context and invoke the model to get a response.

Reference Tutorials Followed

• Tutorial 1: Janakiram MSV
• Tutorial 2: Google Cloud
• Tutorial 3: Google for Developers