Context

Loyal VC is a venture capital firm based out of Toronto, Canada and currently manages a network of 1000+ advisors stored in a database. Advisors are recruited to help portfolio companies with any questions they face during their journey. Loyal incentives advisors by giving them a share of the carry of the fund provided they are helpful to the startup.

To help select the best advisors for each startup, Loyal maintains a database of the advisor’s profiles including Previous Company Experience, Areas of Expertise, and Location data. This information should help Loyal narrow down advisors but unfortunately blends the advisor group together.

Problems with Database Noise 

The Loyal VC database suffers a common problem in organizational databases, bloat and noise. As new data is added, the lack of standardized terms and the introduction of new categories result in database noise. This is a business version of twitter posts which have multiple hashtags. 

Database Noise is the first problem, as by intentional selection, most advisors are good at a lot of areas. When you have two advisors who are experts in 10 areas, it’s hard to distinguish between which one is a true expert and which one has generalist experience. 

Secondly, tags/categories are too simple to capture the nuances in advisor experience. When someone claims to be an expert in M&A transactions is it because their company has been acquired or they have acquired many companies or are they a lawyer who has assisted on hundreds of transactions. All three of these situations are very different to a startup. 

Lastly, Loyal gives portfolio companies access to their database, but it’s still difficult to find the right advisor. A portfolio company might email Loyal:  

“I need help developing a platform which helps to manage acupuncture and other alternative therapies”. 

This is vague and could range from a manager of acupuncture clinics or a request for a programmer who has worked with healthcare data. 

Therefore these vague requests unfortunately require back and forth conversations to better understand the portfolio company needs, slowing down Loyal’s team and the startups. 

THRAG Framework

THRAG, as an approach, is able to solve all of these problems by leveraging vector embeddings, reranking, and large language models (LLMs):

• Building Enriched Profiles: Vector embeddings allow THRAG to create detailed and unique profiles for each advisor. By analyzing the unstructured data from resumes, LinkedIn profiles, and other sources, THRAG can identify and highlight specific skills, experiences, and achievements, going beyond generic tags.

• Nuanced Understanding of Experience: Reranking techniques help THRAG prioritize and order advisors based on the relevance of their experience to the specific needs of a portfolio company. This ensures that advisors with the most pertinent and valuable experience are given precedence, providing a more nuanced and accurate understanding of each advisor's capabilities.

• Natural Language Interactions: With the integration of large language models (LLMs), THRAG allows customers to ask questions in natural language rather than SQL queries. This means portfolio companies can articulate their needs as they would in a conversation, and THRAG can interpret these queries, ask follow-up questions, and refine the search to identify the best advisors efficiently.

The combination of these advanced technologies enables Loyal to overcome the limitations of traditional database approaches, providing a sophisticated and user-friendly solution for matching portfolio companies with the most suitable advisors. This enhanced matching capability ultimately benefits both advisors and portfolio companies, fostering more successful and productive collaborations.