Meet Sallie: Salling Group’s New AI Assistant
In the competitive world of retail, Salling Group has always been at the forefront of innovation. With AI technologies rapidly advancing and influencing our everyday lives, this is no different.
Enter Sallie, the AI assistant designed to enhance and support our colleagues, making their everyday work a little easier.
Introducing Sallie to our colleagues is giving them a new tool to solve their work tasks.
Whether they need help drafting an email, rewriting text, brainstorming creative ideas, or even generating code, Sallie is here to give our colleagues a tool to free up valuable time.
Sallie is a key project within the Generative AI program in Salling Group’s IT&Digital department.
Why Sallie?
We are building Sallie to ensure the assistant is deeply integrated with our specific values, culture, and operational knowledge, providing tailored support that aligns perfectly with our unique needs. Additionally, having a custom-built assistant like Sallie allows for greater control over data privacy and security, ensuring sensitive company information is handled appropriately.
How is Sallie built?
The whole process of building an AI assistant starts with utilizing pre-trained models from known AI services such as OpenAI, Google, and Anthropic.
Sallie’s backend is powered by an LLM proxy which will be able to switch between multiple models. Currently, Sallie utilizes OpenAI’s GPT-4o. The LLM proxy is deployed to Microsoft Azure and runs on a Kubernetes cluster managed by Salling Group, where we can control the data and how it is stored and used. In the future, we might plug and play any other LLM or NLP tools that we consider convenient to enhance the capabilities of the AI assistant.
The LLM proxy is where we implement different features, such as function calling, custom agents, etc. This creates a standardized way of interreacting with the LLM’s across our team.
The LLM proxy API is currently being integrated into other parts of our business, such as our customer service center. Here Sallie is helping the agents to create summaries of a mail correspondence and recommend what the next best action would be.
We are also creating a Salling Group knowledge base that Sallie can access and use to answer questions regarding Salling Group-specific data and information. The knowledge base will consist of data from our various employee handbooks, HR systems, and other sources we can access. Being able to search in a Salling Group database and answer business-relevant questions and provide links to the source takes Sallie to the next level. We are building a search engine infrastructure to retrieve unstructured information in text format from our internal documentation. We want to retrieve the information using queries in natural language and integrate such a search engine with Sallie. The algorithms to build such search engines are rapidly evolving and it is not yet known what the error-proofed way is to compute this information. With this in mind, we have developed a series of tests to benchmark the different parts of the system as we try new approaches. The bare minimum of our search engine is composed of the following modules:
- Index Database: To index the documents, we parse each piece of text into semantically similar sentences. The most basic parsing strategy is to chunk the document in single sentences and a more advanced strategy includes chunking a sequence of tokens using the perplexity score. Once the documents are parsed, each chunk is vectorized using an embedding model. Different embedding models can encode semantic information differently depending on the dataset used to train them. We use a commercial vector database to store and retrieve the information.
- Context Retriever: By storing the transformed documentation from unstructured text to vectors, we create a representational space that can be searched by calculating distances such as the L2. Each query is transformed into a vector using the same embedding model as the database and a number of chunks are retrieved from the vector database.
- Context Evaluator: For each retrieved chunk, a context evaluator is used to ensure that the retrieved contextual information is semantically relevant to answer the question.
- Question Answering: With a list of contextually relevant information, we generate an answer that can be integrated into the chat history in Sallie.
What can Sallie do?
But enough talk, let’s have a look at some current use cases where Sallie can help.
Firstly, looking at job postings.
We see clear benefits in using Sallie to check job postings for bias and to improve the wording.
The second example we want to show is how handy Sallie can be when you need someone to brainstorm ideas with. Sometimes you need inspiration for an important presentation or how to solve a complex task. Sallie is right there by your side, providing you with the help you need.
Conclusion
Sallie represents a significant step forward in how Salling Group leverages AI to support our colleagues and enhance our operations. By integrating deeply with our values and culture, Sallie ensures that the support provided is both relevant and secure. As we continue to develop and refine Sallie, we look forward to seeing the positive impact it will have across our organization. Stay tuned for more updates as we continue to innovate and improve everyday life for our colleagues and customers.
My name is Simon Banerjee, and I am an Advanced Data Scientist in the ML/AI team at Salling Group. Here, we leverage machine learning and artificial intelligence to drive innovation and enhance our data-driven decision-making processes. I am passionate about solving business problems and continuously developing new methodologies and processes. I believe I make a significant impact at Salling Group by advancing our technological capabilities. In my work, I strive to bring innovation, focusing on improving and optimizing solutions to better serve our customers and colleagues.
