qabot 0.3.2

Query local or remote data files with natural language queries powered by OpenAI and DuckDB.

qabot

Query local or remote files with natural language queries powered by langchain and gpt and duckdb 🦆.

Can query Wikidata and local files.

Command Line Usage

$ EXPORT OPENAI_API_KEY=sk-...
$ EXPORT QABOT_MODEL_NAME=gpt-4
$ qabot -w -q "How many Hospitals are there located in Beijing"
Query: How many Hospitals are there located in Beijing
There are 39 hospitals located in Beijing.
Total tokens 1749 approximate cost in USD: 0.05562

Python Usage

from qabot import ask_wikidata, ask_file

print(ask_wikidata("How many hospitals are there in New Zealand?"))
print(ask_file("How many men were aboard the titanic?", 'data/titanic.csv'))

Output:

There are 54 hospitals in New Zealand.
There were 577 male passengers on the Titanic.

Features

Works on local CSV files:

remote CSV files:

$ qabot \
    -f https://www.stats.govt.nz/assets/Uploads/Environmental-economic-accounts/Environmental-economic-accounts-data-to-2020/renewable-energy-stock-account-2007-2020-csv.csv \
    -q "How many Gigawatt hours of generation was there for Solar resources in 2015 through to 2020?"

Even on (public) data stored in S3:

You can even load data from disk via the natural language query, but that doesn't always work...

"Load the file 'data/titanic_survival.parquet' into a table called 'raw_passengers'. Create a view of the raw passengers table for just the male passengers. What was the average fare for surviving male passengers?"

After a bit of back and forth with the model, it gets there:

The average fare for surviving male passengers from the 'male_passengers' view where the passenger survived is 40.82. I ran the query: SELECT AVG(Fare) FROM male_passengers WHERE Survived = 1 AND Sex = 'male'; The average fare for surviving male passengers is 40.82.

Quickstart

You need to set the OPENAI_API_KEY environment variable to your OpenAI API key, which you can get from here.

Install the qabot command line tool using pip/poetry:

$ pip install qabot

Then run the qabot command with either local files (-f my-file.csv) or -w to query wikidata.

Examples

Local CSV file/s

$ qabot -q "how many passengers survived by gender?" -f data/titanic.csv
🦆 Loading data from files...
Loading data/titanic.csv into table titanic...

Query: how many passengers survived by gender?
Result:
There were 233 female passengers and 109 male passengers who survived.


 🚀 any further questions? [y/n] (y): y

 🚀 Query: what was the largest family who did not survive? 
Query: what was the largest family who did not survive?
Result:
The largest family who did not survive was the Sage family, with 8 members.

 🚀 any further questions? [y/n] (y): n

Query WikiData

Use the -w flag to query wikidata. For best results use the gpt-4 model.

$ EXPORT QABOT_MODEL_NAME=gpt-4
$ qabot -w -q "How many Hospitals are there located in Beijing"

Intermediate steps and database queries

Use the -v flag to see the intermediate steps and database queries. Sometimes it takes a long route to get to the answer, but it's interesting to see how it gets there.

qabot -f data/titanic.csv -q "how many passengers survived by gender?" -v

Data accessed via http/s3

Use the -f <url> flag to load data from a url, e.g. a csv file on s3:

$ qabot -f s3://covid19-lake/enigma-jhu-timeseries/csv/jhu_csse_covid_19_timeseries_merged.csv -q "how many confirmed cases of covid are there?" -v
🦆 Loading data from files...
create table jhu_csse_covid_19_timeseries_merged as select * from 's3://covid19-lake/enigma-jhu-timeseries/csv/jhu_csse_covid_19_timeseries_merged.csv';

Result:
264308334 confirmed cases

Links

• Python library docs
• Agent docs to talk to arbitrary apis via OpenAPI/Swagger
• Agents/Tools to talk SQL
• Typescript library

Ideas

• Decent Python Library API so can be used from other Python code
• streaming mode to output results as they come in
• token limits
• Supervisor agent - assess whether a query is "safe" to run, could ask for user confirmation to run anything that gets flagged.
• Often we can zero-shot the question and get a single query out - perhaps we try this before the MKL chain
• test each zeroshot agent individually
• Generate and pass back assumptions made to the user
• Add an optional "clarify" tool to the chain that asks the user to clarify the question
• Create a query checker tool that checks if the query looks valid and/or safe
• Inject AWS credentials into duckdb so we can access private resources in S3
• Better caching