Large Language Models in Finance

Frequently Asked Questions

• Do you need to train your own LLM from scratch to stay competitive?

  While there are companies who have gone the route of training their own LLM from scratch - spending millions of dollars in the process - it is actually a better idea to remain flexible if you want to stay competitive. Vendor agnosticism, a principle championed by deepset, allows you to change models when they no longer serve you. For example, if a cheaper or faster model comes along, you can simply plug it into your existing pipeline and move on. Fine-tuning an existing model, on the other hand, can be a viable solution for adapting an LLM to the specific language of the financial domain. We're happy to share our years of experience in customizing models for different applications to help you build the solution that's perfect for your use case.

• How can you keep your data safe?

  In the age of AI models and distributed computing infrastructures, data security is a major concern - even more so in an area as sensitive as finance. At deepset, we understand this and have made data security a priority. Users can manage access with MFA and SSO in deepset Cloud, while our Virtual Private Cloud (VPC) option gives them the flexibility to keep their data layer in their preferred location. We are also SOC 2 and GDPR compliant and CSA STAR Level 1 certified.

• Who do I need to hire if I want to start building with AI?

  To build products or internal tools with AI, you need to put together a team that understands AI technology, has a product mindset, and understands both user needs and business requirements. This type of team is called an "AI team." It can be large or small, as long as it has the right cross-functional skills. To learn more about AI teams and how to build one, check out our whitepaper on Leading a successful AI team.

• My stakeholders are concerned about LLM hallucinations. How can I reassure them?

  LLMs "hallucinate", that is, they make up facts that are not supported by any data. Because of the eloquence of these models, hallucinations can be difficult to detect, creating a volatile factor that is a barrier to using LLMs in production. However, using a combination of prevention techniques, teams can reduce the number of hallucinations to a minimum. These include effective prompting, grounding the LLM's response in fact-checked data through retrieval augmented generation (RAG), and monitoring the Groundedness of responses.