Avoiding LLM hallucinations through analytical AI, is it possible?

The success of ChatGPT has generated massive interest in generative AI, especially in the language processing field. LLM models are rapidly improving, and there are increased initiatives to use them within companies.

For customer support, automating the processing of incoming requests is a popular topic. For example, automatically generating draft responses can have several advantages: it saves time in processing requests and ensures consistency in responses across agents.

However, generative AI can produce false information, known as hallucinations, which can have critical consequences. Spreading incorrect information can lead to customer dissatisfaction and impact the company’s reputation.

On the other hand, analytical AI, such as the one developed at Golem.ai and used in InboxCare, does not suffer from hallucinations.

Framework

Our study tested combinations of message information (subject, customer signature) with InboxCare data (category, order status, customer name) using the Llama 2 70b model on H100 GPUs hosted by Scaleway.

Key Findings

• LLM is already capable of understanding information in messages without InboxCare data

• Order status information from InboxCare significantly improves result quality

• Prompt engineering greatly influences response quality and accuracy

• Response size limits (40 words) and minimum temperature (0.01) most effectively reduce hallucinations

• Paradox: Limiting hallucinations requires strict constraints that reduce creativity

Conclusions

• Adding redundant information from InboxCare doesn’t improve results

• New/unique data from InboxCare (order status) improves quality

• Question arises: Is LLM worth using vs. template-based systems?

• LLMs cannot guarantee complete elimination of hallucinations

Recommended Next Steps

• Explore other LLM models (e.g., Mistral 7b)

• Refine prompt engineering techniques

• Fine-tune LLM for customer support use cases

• Trigger LLM use only for specific message categories based on acceptable hallucination risk