While most retrieval augmented LLMs do have the functionality to include citations on their answers, citations donot reveal anything about the numerical, contextual or logical correctness of the LLM’s answer especially when reconciling information from 2 or more sources.
Most generic RAG implementations provide citations to assure users of the “groundedness” of the answer the LLM has generated. However, citations alone cannot ensure the accuracy or coherence of LLM answers, especially when integrating information from multiple sources. Hence, additional processing is needed to address the numerical, contextual, and logical correctness of the answers.
<aside> 💡 Source data reconciliation is crucial in financial services as it tackles important metrics such as addressing bias, evaluating credibility, and ensuring the relevance of multiple data sources. Hence, verifying data provenance becomes important here.
</aside>
It is important to verify that all the numerical data mentioned in the LLM’s answers is either present in one of the documents retrieved to answer that question or the number is directly derivable from the formula of any financial term explicitly specified in the question by a user. Building this number lineage is especially important because LLMs do not treat numbers differently from text data. The most common error which can slip by unnoticed here is the misplacement of decimal points.
Fig 1: Concept of how we currently show users answer grounding
Even if an LLM has parsed the numbers correctly from the original sources, it can still misrepresent the context of a particular sentence in its answer. For example: instead of saying that profits have increased by 69%, it can say that the current net profit margin is 69%, so while the number is right the context in which it has been represented is incorrect. This can trip up unsuspecting research analysts who might just believe that because the number mentioned is correct, so is its context.
<aside> 🤔 One easy way to visualize LLMs as not being AGI is that they cannot comprehend logical symbols any differently than a sequence of tokens. This means that giving it 100 books on logical reasoning isn't enough to make it logical, otherwise only 1 book would suffice.
</aside>
The logical correctness of an answer is crucial because even if the context and numbers are accurately derived by an LLM, it can still misrepresent logic. For example, stating that the net profit is 69% instead of mentioning that profits increased by 69% YoY due to a 10% increase in operating margin can mislead an analyst’s sense of correlation and possible causation.