Kinetica SqlAssist LLM Demo
This notebook demonstrates how to use Kinetica to transform natural language into SQL and simplify the process of data retrieval. This demo is intended to show the mechanics of creating and using a chain as opposed to the capabilities of the LLM.
Overview
With the Kinetica LLM workflow you create an LLM context in the database that provides
information needed for infefencing that includes tables, annotations, rules, and
samples. Invoking ChatKinetica.load_messages_from_context() will retrieve the
context information from the database so that it can be used to create a chat prompt.
The chat prompt consists of a SystemMessage and pairs of
HumanMessage/AIMessage that contain the samples which are question/SQL
pairs. You can append pairs samples to this list but it is not intended to
facilitate a typical natural language conversation.
When you create a chain from the chat prompt and execute it, the Kinetica LLM will
generate SQL from the input. Optionally you can use KineticaSqlOutputParser to
execute the SQL and return the result as a dataframe.
Currently, 2 LLM's are supported for SQL generation:
- Kinetica SQL-GPT: This LLM is based on OpenAI ChatGPT API.
- Kinetica SqlAssist: This LLM is purpose built to integrate with the Kinetica database and it can run in a secure customer premise.
For this demo we will be using SqlAssist. See the Kinetica Documentation site for more information.
Prerequisites
To get started you will need a Kinetica DB instance. If you don't have one you can obtain a free development instance.
You will need to install the following packages...
# Install Langchain community and core packages
%pip install --upgrade --quiet langchain-core langchain-community
# Install Kineitca DB connection package
%pip install --upgrade --quiet 'gpudb>=7.2.0.8' typeguard pandas tqdm
# Install packages needed for this tutorial
%pip install --upgrade --quiet faker ipykernel
Database Connection
You must set the database connection in the following environment variables. If you are using a virtual environment you can set them in the .env file of the project:
KINETICA_URL: Database connection URLKINETICA_USER: Database userKINETICA_PASSWD: Secure password.
If you can create an instance of KineticaChatLLM then you are successfully connected.
from langchain_community.chat_models.kinetica import ChatKinetica
kinetica_llm = ChatKinetica()
# Test table we will create
table_name = "demo.user_profiles"
# LLM Context we will create
kinetica_ctx = "demo.test_llm_ctx"
API Reference:
Create test data
Before we can generate SQL we will need to create a Kinetica table and an LLM context that can inference the table.
Create some fake user profiles
We will use the faker package to create a dataframe with 100 fake profiles.
from typing import Generator
import pandas as pd
from faker import Faker
Faker.seed(5467)
faker = Faker(locale="en-US")
def profile_gen(count: int) -> Generator:
for id in range(0, count):
rec = dict(id=id, **faker.simple_profile())
rec["birthdate"] = pd.Timestamp(rec["birthdate"])
yield rec
load_df = pd.DataFrame.from_records(data=profile_gen(100), index="id")
print(load_df.head())
username name sex \
id
0 eduardo69 Haley Beck F
1 lbarrera Joshua Stephens M
2 bburton Paula Kaiser F
3 melissa49 Wendy Reese F
4 melissacarter Manuel Rios M
address mail \
id
0 59836 Carla Causeway Suite 939\nPort Eugene, I... meltondenise@yahoo.com
1 3108 Christina Forges\nPort Timothychester, KY... erica80@hotmail.com
2 Unit 7405 Box 3052\nDPO AE 09858 timothypotts@gmail.com
3 6408 Christopher Hill Apt. 459\nNew Benjamin, ... dadams@gmail.com
4 2241 Bell Gardens Suite 723\nScottside, CA 38463 williamayala@gmail.com
birthdate
id
0 1997-12-08
1 1924-08-03
2 1933-12-05
3 1988-10-26
4 1931-03-19
Create a Kinetica table from the Dataframe
from gpudb import GPUdbTable
gpudb_table = GPUdbTable.from_df(
load_df,
db=kinetica_llm.kdbc,
table_name=table_name,
clear_table=True,
load_data=True,
)
# See the Kinetica column types
print(gpudb_table.type_as_df())
name type properties
0 username string [char32]
1 name string [char32]
2 sex string [char2]
3 address string [char64]
4 mail string [char32]
5 birthdate long [timestamp]
Create the LLM context
You can create an LLM Context using the Kinetica Workbench UI or you can manually create it with the CREATE OR REPLACE CONTEXT syntax.
Here we create a context from the SQL syntax referencing the table we created.
from gpudb import GPUdbSamplesClause, GPUdbSqlContext, GPUdbTableClause
table_ctx = GPUdbTableClause(table=table_name, comment="Contains user profiles.")
samples_ctx = GPUdbSamplesClause(
samples=[
(
"How many male users are there?",
f"""
select count(1) as num_users
from {table_name}
where sex = 'M';
""",
)
]
)
context_sql = GPUdbSqlContext(
name=kinetica_ctx, tables=[table_ctx], samples=samples_ctx
).build_sql()
print(context_sql)
count_affected = kinetica_llm.kdbc.execute(context_sql)
count_affected
CREATE OR REPLACE CONTEXT "demo"."test_llm_ctx" (
TABLE = "demo"."user_profiles",
COMMENT = 'Contains user profiles.'
),
(
SAMPLES = (
'How many male users are there?' = 'select count(1) as num_users
from demo.user_profiles
where sex = ''M'';' )
)
1
Use Langchain for inferencing
In the example below we will create a chain from the previously created table and LLM context. This chain will generate SQL and return the resulting data as a dataframe.
Load the chat prompt from the Kinetica DB
The load_messages_from_context() function will retrieve a context from the DB and convert it into a list of chat messages that we use to create a ChatPromptTemplate.
from langchain_core.prompts import ChatPromptTemplate
# load the context from the database
ctx_messages = kinetica_llm.load_messages_from_context(kinetica_ctx)
# Add the input prompt. This is where input question will be substituted.
ctx_messages.append(("human", "{input}"))
# Create the prompt template.
prompt_template = ChatPromptTemplate.from_messages(ctx_messages)
prompt_template.pretty_print()
API Reference:
================================[1m System Message [0m================================
CREATE TABLE demo.user_profiles AS
(
username VARCHAR (32) NOT NULL,
name VARCHAR (32) NOT NULL,
sex VARCHAR (2) NOT NULL,
address VARCHAR (64) NOT NULL,
mail VARCHAR (32) NOT NULL,
birthdate TIMESTAMP NOT NULL
);
COMMENT ON TABLE demo.user_profiles IS 'Contains user profiles.';
================================[1m Human Message [0m=================================
How many male users are there?
==================================[1m Ai Message [0m==================================
select count(1) as num_users
from demo.user_profiles
where sex = 'M';
================================[1m Human Message [0m=================================
[33;1m[1;3m{input}[0m
Create the chain
The last element of this chain is KineticaSqlOutputParser that will execute the SQL and return a dataframe. This is optional and if we left it out then only SQL would be returned.
from langchain_community.chat_models.kinetica import (
KineticaSqlOutputParser,
KineticaSqlResponse,
)
chain = prompt_template | kinetica_llm | KineticaSqlOutputParser(kdbc=kinetica_llm.kdbc)
API Reference:
Generate the SQL
The chain we created will take a question as input and return a KineticaSqlResponse containing the generated SQL and data. The question must be relevant to the to LLM context we used to create the prompt.
# Here you must ask a question relevant to the LLM context provided in the prompt template.
response: KineticaSqlResponse = chain.invoke(
{"input": "What are the female users ordered by username?"}
)
print(f"SQL: {response.sql}")
print(response.dataframe.head())
SQL: SELECT username, name
FROM demo.user_profiles
WHERE sex = 'F'
ORDER BY username;
username name
0 alexander40 Tina Ramirez
1 bburton Paula Kaiser
2 brian12 Stefanie Williams
3 brownanna Jennifer Rowe
4 carl19 Amanda Potts