Swiggy has developed an in-house text-to-SQL AI tool, Hermes, that allows its workers to generate and execute SQL queries via natural language prompts in Slack.
In an effort to improve data access and decision-making speeds for its teams, Swiggy officially launched Hermes to streamline operations and improve efficiency, making use of a generative AI.
“Hermes is our in-house developed generative AI-based workflow that allows a user to ask a question in natural language and a) get a SQL query generated, and b) receive results by automatically executing the generated query, all within Slack,” the company said in a blog post.
Hermes is currently on its second iteration, with Swiggy stating that Hermes V1 was built closer to already existing enterprise solutions, with “minimal Swiggy-specific modifications.”
Based on their learnings from V1, the company has now launched V2, which has compartmentalised different business units, called charters. Interestingly, each charter has its own metadata and is tuned to address specific use cases for their relevant teams. V2 now makes use of Slack, AWS Lambda, Databricks and GPT-4o in generating inputs for its users.
Hermes V2 allows users to type in their queries into Slack, following which it will go through AWS Lambda, which acts as an intermediary between the user input and the GenAI model, to process and reformat the user input. Following this, a Databricks job is created, which will fetch the relevant charter’s AI model, generate an SQL query based on the input and execute it.
Additionally, the company has created a knowledge base and makes use of RAG to ensure Swiggy-specific context, allowing GPT-4o to choose the relevant tables and columns needed.
“Hundreds of users across the company have been using Hermes over the last few months to answer several thousand queries with an average turnaround time of <2 minutes.
“This approach also streamlined/modularised the product-lifecycle management, enabling us to onboard new charters, test outputs, and make continuous adjustments to the knowledge base as needed,” they stated.
In the future, the company plans on adding a knowledge base of historical queries, a query explanation layer and a ReAct agent to the tool to improve overall accuracy and provide its teams with a better idea of how it works.
This is only the latest in GenAI developments within the company, as Swiggy has been pretty open to the use of AI within its workings, thanks to the establishment of a GenAI task force at the beginning of 2023. Since then, the company has made use of GenAI for partner support as well as generating images for the platform itself.
