This project demonstrates a simple end-to-end Retrieval-Augmented Generation (RAG) application on Databricks (tested on Serverless compute) using Large Language Models (LLMs), embedding model, Databricks-native vector search APIs, LangChain, MLflow, Streamlit, etc. It uses Streamlit as frontend / graphical user interface for question answering over your own document if the document can be found in vector database, else it will get the answer from the Internet.
-
Fork this repository to your repository in Databricks or any IDE tools.
-
You must provide your own Databricks personal access token and workspace URL in app.yaml file (never share your personal access token publicly).
-
You must execute all cells in rag_app notebook file in order to create relevant tables and vector_search_endpoint.
-
You can adjust below variables in rag_app notebook file and app.py file (applicable to some variables) to suit your environment.
- llm_model = "databricks-gemma-3-12b"
- embedding_model = "databricks-bge-large-en"
- vector_search_endpoint = "test_vector_search"
- catalog_name = "workspace"
- schema_name = "my_schema"
- volume_name = "my_volume"
- table_doc_text = f"{catalog_name}.{schema_name}.rag_doc_text"
- table_doc_page = f"{catalog_name}.{schema_name}.rag_doc_page"
- table_doc_chunk = f"{catalog_name}.{schema_name}.rag_doc_chunk"
- table_doc_embedding = f"{catalog_name}.{schema_name}.rag_doc_embedding"
-
You can modify rag_app notebook file if you need to change the data/document ingestion pipeline, chunking strategy, embedding model, etc. Basically this is the core file for RAG development work.
-
Below 3 files are required for application deployment in Databricks.
- app.py
- app.yaml
- requirements.txt
-
If you need to install additionally library, you can put it in requirements.txt file together with the library version. During deployment, all required libraries will be installed based on requirements.txt file.
-
Once you have successfully run all cells in rag_app notebook file, you will see below vector search endpoint named test_vector_search.
-
-
Deployment steps
- Click on Compute on left panel / sidebar, then go to Apps tab.
- In the Apps tab, click Create app button.
- In Create new app screen, select Create a custom app.
- In Create new app screen, for App name field, enter rag, then click the Create app button.
- It will take a while to create the application instance.
- Once the application instance is created, then click Deploy button.
- In Create deployment window, select the workspace location when you keep all of the 4 files above and click Deploy button to continue.
- Once deployed successfully, then you can see below screen.
- You can copy the newly generated URL link and paste it in your browser to test it. Alternatively you can launch it by clicking the icon after the URL link.
- Below screen shows the answer from vector database.
- Below screen shows the answer from Internet (with a WARNING label) as it cannot retrieve the answer from vector database.
- Databricks will start and run the Streamlit application (using command: streamlit run app.py) automatically.
