1 of 19

AI based Simplified Database Management�

Suhas Basavaraj

Lead Consultant

2 of 19

Agenda

Introduction
Concepts
Use case
POC Component
Architecture
Demo
Next steps

3 of 19

Introduction

Manually executing database administration tasks , analysing operational issues and optimizing the performance can negatively impact the operational efficiency , reduce the employee productivity

Gen AI solution for the POC is done will address below use cases

Provide the guidance for database administration
Recommend best practices for the database health check
Recommend query tuning best practices

4 of 19

Concepts

What is a vector ?

In simple terms, a vector is a list of numbers that can represent a point in space. For example, in 2D space, a vector [2,3] represents a point that’s 2 units along the x-axis and 3 units along the y-axis.

What is Vector embedding ?

Vector embeddings allow us to convert diverse forms of data into a common format (vectors) that LLMs can understand and process. By doing so, we can perform mathematical operations on them.

5 of 19

Vector embeddings

Cricket:[−0.7888,−0.7361,−0.6208,−0.5134,−0.4044]

Cricket: [0.8108,0.6671,0.5565,0.5449,0.4466]

Grasshopper: [0.8308,0.6805,0.5598,0.5184,0.3940]

Baseball: [−0.7715,−0.7300,−0.5986,−0.4908,−0.4454]

Graphical representation

Vector in 2D representation

6 of 19

PostgreSQL as vector database

Pgvector is an extension for PostgreSQL

CREATE EXTENSION IF NOT EXISTS vector;

ACID, point-in-time recovery, partitioning, scalability
Vector datatype
Exact (Default) and approximate nearest neighbour search, L2 distance, inner product, and cosine distance
pgvector along with our business data in the same modern database
Pgvector also supports indexing-Inverted File (IVFFlat) and Hierarchical Navigable Small Worlds (HNSW)

7 of 19

8 of 19

Large Language Models

System that uses deep learning techniques to generate human-like text based on patterns learned

Vast datasets are used for pretraining

Self-supervised learning techniques (e.g., masked language modelling, next sentence prediction)

Example: anthropic.claude-3-sonnet-20240229-v1:0, amazon.titan-embed-text-v1

9 of 19

Prompt engineering

Input text or instruction given to a language model to generate a desired output

Context: Context refers to the additional information or background provided to LLM

Optimizations –

One-shot Prompting
Few-shot Prompting
Chain-of-Thought Prompting

10 of 19

Retrieval-Augmented Generation (RAG)

Enhancing or supplementing the language model's capabilities by incorporating external knowledge or information
Providing this retrieved knowledge to the language model, in addition to the input prompt or context
The language model then generates text while considering both the input and the augmented retrieved knowledge.

11 of 19

Flow Diagram

12 of 19

POC Use Case

Objective Business Value Outcome

Reduce the database management cost by offloading the database administration tasks

Remove the need for users to learn complex database tasks like database tuning and troubleshooting

Improve the operational efficiency by quickly providing the tuning and troubleshooting based on the customer specific data and tested knowledgebase

�

Database specific guidance for administrative tasks , troubleshooting and query tuning

Higher accuracy and fast response as its generates the output from customer specific data

13 of 19

POC -Scope

Developed the earlier version of Database Management Assistant which mimic human and provides the guidance for database queries

Use RAG , Knowledge base and LLM to build the solution

Use the real time data to fetch the metadata from the database

Use Langchain for session level contextual information

Use this to framework to build, test and improve

14 of 19

Tools

Python
�LLM
�Streamlit
�Langchain

15 of 19

Architecture

16 of 19

Steps

Database Management Assistant

Create the internal RAG knowledge base for user queries
Prompt the user for the required information
Save the conversational contextual information using the lang chain
Query the Foundation model for the context

B. Recommend the implementation best practices and tuning recommendation

Create the internal RAG knowledge base based on execution plan and recommendation
Prompt the user for query details
Use the RAG to build the context and retrieve the information from the knowledge base.
Query the Foundation model for the context

17 of 19

Demo

Demo and code walkthrough

18 of 19

Next steps

Troubleshooting and recommendation

Provide the commands based on real time customer metadata

Knowledge base enrichment for higher accuracy

Incorporate the conversational history for context building

Leveraging the chain of thoughts

Implement the data governance

19 of 19

Thank you!