Overview
Our experts have been developing a demo of an intelligent search agent, a tool that uses artificial intelligence for information retrieval. In our demo, we use large language models to process a user’s search query. From there, the intelligent search agent answers a user’s questions by interfacing with multiple sources to generate relevant and accurate responses.
For our large language model, we use GPT-4 with OpenAI’s API. Used on its own, ChatGPT has a couple of limitations. First, the model behind ChatGPT is trained on data through 2021, so a user cannot receive the most up-to-date information. Second, at times, the chatbot can be prone to hallucination, a phenomenon in which the chatbot makes incorrect claims not directly based on material in their training sets.
To improve the accuracy and reliability of our search agent, we connect GPT-4 to additional tools. More specifically, we supply a list of sources, like Google and Wikipedia, from which the search agent can draw information relevant to the user input. In principle, however, any news source could also be connected. We also use prompt engineering to guide the language model and improve the quality of the responses.
After the user asks the search agent a question about a company, the user receives a response to those questions. The response includes references to the sourced data, thereby enhancing the reliability of the response.
Use Case
As a concrete example, suppose you are curious about how an organization manages its risks and opportunities around sustainability issues. One proxy for that is an ESG score, which is a measure of how a company fares with respect to environmental, social, and governance-related issues, such as carbon emissions, employee diversity, and data security. Investors can use an ESG score to analyze their potential future investments, whereas organizations can use their own ESG score as a tool to evaluate performance and improve decision-making.
Both internal and external stakeholders generate ESG scores. From an external viewpoint, generating an ESG score can involve a time-consuming process of analyzing corporate disclosures, conducting management interviews, and reviewing public information about an organization.
To streamline this process, you may use the search agent to ask: “Does {company} produce oil and gas?” as well as “Is {company} involved in the oil and gas industry?”.
In response to each question, the agent provides a simple “yes” or “no” with an elaboration below the response. For a snapshot of the demo, we choose a consulting company that happens to provide services to the oil and gas industry. Despite the subtle difference between these two questions, the search agent processes the user’s query accurately.
Additionally, the elaboration contains links to various resources that can help the user explore the responses on a deeper level. We see that this agent can be extremely helpful in conducting research on how companies approach specific issues related to the ESG score. Improving the research methods, in turn, helps facilitate a more objective ESG scoring procedure.
Overall, through this example, we see how large language models, when combined with prompt engineering, can understand nuance in the user’s queries.
Closing thoughts
As the landscape surrounding large language models continues to evolve rapidly, our technical team eagerly develops LLM-oriented solutions with the latest tools available. Our search agent demo conveys that we strive to develop elegant solutions that have interesting and impactful applications.
Resources:
[1] ESG Scoring: ESG Score – Definition, Process, Implications & Purpose (corporatefinanceinstitute.com)
