This project intends to provide a quick way to integrate PaLM with BigQuery in Python, using SQLAlchemy and LangChain to glue together the pieces. Added Streamlit for a WebUI. This is in no way a production ready artefact, it rather provides a proof of concept for testing purposes.
The inception of this project was to evaluate if BigQuery could be interrogated using an LLM. The results were rather satisfying, hence the publication. While not provided here, the integration with ChatGPT worked as well and can be switched with PaLM easily.
PaLM 2 is the latest LLM provided by Google as a cloud service that is quite performant and easy to access through API using Python. Note that at the moment writing this, PaLM is not widely available in Europe and you may need to access it through a server/services in the US.
BigQuery is a powerful managed DataWarehouse that provides an out of the box and serverless analytical database. Users often use it to store vast amounts of Data. In this example, we have used the following public and free Thelook Dataset.
LangChain provides an OSS suite of building blocks to chain LLM oriented tasks with reasoning like database queries, math, document loaders, vector dbs or parsers.
SQL Alchemy is a Python toolkit to provide an abstract layer to many popular databases.
(Optional)Streamlit is a lighweight WebUI wrapper for Python.
In order to launch the script you will need the following :
- A working Python 3 environment
- Install dependencies using pip
- A Google PaLM API key
- A service account on Google Cloud with BigQuery Data Viewerrights and download the JSON key file.
- BigQuery API needs to be activated.
To install the dependencies, you will need to run the following command
pip install langchain chromadb tiktoken tabulate sqlalchemy sqlalchemy-bigquery google-cloud-bigquery google-generativeai google-cloud-bigquery-storage streamlit
The following variables need to be set in the launcher-PaLM.py script before execution
service_account_file = "your_sa_account_key_file.json"
os.environ['GOOGLE_API_KEY'] = 'your_palm_api_key'
project = "your_project_id"
dataset = "your_dataset" # Defined when you imported TheLook Dataset
table = "you_table"
Notes The JSON service account key file should be accessible to your script.
The variable your_project_id can be found here.
The variable your_dataset is defined when you import TheLook dataset from AnalyticsHub. You can set your_table to the events table from TheLook. You can also comment out the table variable as SQL Alchemy car handle the whole schema. This will allow to JOIN tables but can generate more hallucinations or false/positives.
You can run the script using the following command. This will will run in the console and exit.
python3 launcher-PaLM.py
Example output
> Entering new AgentExecutor chain...
Action: sql_db_list_tables
Action Input:
Observation: events
Thought:The "events" table seems like it could contain information about user events, which might include data on publisher sources. I need to check the schema of the "events" table to find the relevant columns that could contain publisher source information.
Action: sql_db_schema
Action Input: events
Observation:
CREATE TABLE `events` (
`user_id` INT64,
`sequence_number` INT64,
`session_id` STRING,
`created_at` TIMESTAMP,
`ip_address` STRING,
`city` STRING,
`state` STRING,
`postal_code` STRING,
`browser` STRING,
`uri` STRING,
`event_type` STRING,
`ide` INT64,
`traffic_sources` STRING
)
/*
3 rows from events table:
user_id sequence_number session_id created_at ip_address city state postal_code browser uri event_type ide traffic_sources
41525 4 c75439e0-127d-4721-be6c-b03af9255b3f 2020-10-31 16:57:38+00:00 118.90.172.146 Bogatynia Dolnośląskie 59 Chrome /cart cart 539126 Email
None 3 2fc7f38e-d997-4c26-b493-48ad47aaf8b0 2022-04-07 02:35:00+00:00 83.132.130.52 Bogatynia Dolnośląskie 59 Safari /cart cart 2271896 Adwords
None 3 e35a5c16-85e2-412a-8226-5fe9285ab994 2021-12-29 03:09:00+00:00 162.176.5.118 Bogatynia Dolnośląskie 59 Chrome /cart cart 2148147 Email
*/
Thought:The `traffic_sources` column in the `events` table seems to be the relevant column for publisher sources. I will write a query to select the top 3 traffic sources by the count of their occurrences.
Action: sql_db_query_checker
Action Input: SELECT traffic_sources, COUNT(*) as source_count FROM `events` GROUP BY traffic_sources ORDER BY source_count DESC LIMIT 3
Observation: The provided query does not seem to contain any of the common mistakes listed. It is a simple query that selects the `traffic_sources` column from the `events` table, counts the number of occurrences of each source, groups the results by `traffic_sources`, orders them in descending order by the count, and limits the output to the top 3 sources.
Here is the original query:
sql
SELECT traffic_sources, COUNT(*) as source_count FROM `events` GROUP BY traffic_sources ORDER BY source_count DESC LIMIT 3
Thought:The query checker has confirmed that the query does not contain any common mistakes. Now I will execute the query to get the top 3 publisher sources.
Action: sql_db_query
Action Input: SELECT traffic_sources, COUNT(*) as source_count FROM `events` GROUP BY traffic_sources ORDER BY source_count DESC LIMIT 3
Observation: [('Email', 1092153), ('Adwords', 730558), ('Facebook', 243532)]
Thought:I now know the final answer
Final Answer: The top 3 publisher sources from which users were coming are Email, Adwords, and Facebook.
> Finished chain.
Using this streamlit implementation assumes you have a Graphic UI on the system you are exectuting the script on so you can have a browser to access the WebUI. This will not work with a remotely executed python notebook for instance (e.g. Google Cloud Workbench)
Change the agent's execution from this
agent_executor.run("From what top 3 publisher sources were users coming from?")
'''
if prompt := st.chat_input():
st.chat_message("user").write(prompt)
with st.chat_message("assistant"):
st_callback = StreamlitCallbackHandler(st.container())
response = agent_executor.run(prompt, callbacks=[st_callback])
st.write(response)
'''
To this
'''
agent_executor.run("From what top 3 publisher sources were users coming from?")
'''
if prompt := st.chat_input():
st.chat_message("user").write(prompt)
with st.chat_message("assistant"):
st_callback = StreamlitCallbackHandler(st.container())
response = agent_executor.run(prompt, callbacks=[st_callback])
st.write(response)
You can then run streamlit the following command. This will will run in the console and open an browser locally to provide the WebUI.
streamlit run launcher-PaLM.py
Example output
One great features of SQL Alchemy with LangChain is that you do not need to setup a schema as it is created on the fly.
The temperature can be adjusted to throttle the "creativity" of your LLM. The top_k can be adjusted according to the token size accepted by your API.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent