HR Slack Bot

This is a Slack app using the Python slack-sdk to extract emails from a certain list of channels and parse them to a structured schema and push the data to a Google Spreadsheet. The bot tries to have the same functionality as other bots using slash commands.

What's happening under the hood?

As a email body gets posted from one of the observed Slack channels, the bot uses the Slack SDK to retrieve the messages as a JSON string and push them to a SQLite database stored locally (ofc) We use Slack's rounded timestamp ( ts) as the primary key for each message, so we are only adding the messages that are new to the messsages table.

Emails are parsed using gpt-3-1102 with no temperature to avoid generation and the data is retrieved using a data schema (see src/schemas.py). The data is taken from the table messages in the SQL database (text column, following the slack-sdk standard). Once processed, the parsed data is stored in the parsed_messages table with the ts identifier and the channel_id, and later pushed to a Google Spreadsheet.

Language model configuration

• Prompting design happens in src/extractor.py, but changes to the prompt are done in other parts of the script. We follow the following prompting strategy:

main_prompt -> examples -> response_schema 

• Examples are a way to retrieve more accurate responses for data that the model hasn't seen. We use the examples as a way to improve retrieval and we can define better examples in data/examples.json Notice any example should include a text prompt and the outcome we want.