Page 19 - AI Vol 1: Foundations of AI
P. 19
of LLMs and is deployed in many popular LLM solution. As previously noted, LLMs can still
products, such as ChatGPT, Microsoft Copilot, produce hallucinations even when referencing
and PerplexityAI. The basic principle is that information contained in their context windows.
LLMs produce more accurate results when Secondly, the output will be heavily influenced
they have information available to them in their by the quality of the information retrieved, which
context window. For example, if a public agency itself is influenced by how the LLM conducts
employee asks an LLM a question about some its Internet or database search. A key challenge
issue facing the agency, the LLM would produce in retrieval-augmented generation is enabling
an answer based on the information and patterns LLMs to assess the relevance and reliability
it learned in its training. However, if the user of the information they retrieve. Finally, as
incorporated relevant policies, regulations, and previously discussed, the limitations of the size
other information into their prompt, the LLM of the context window and the functioning of the
would analyze this additional information and attention mechanism mean that the LLM will be
refer to it in its response, effectively grounding limited in how much factual information it can
the LLM in a source of “truth” and reducing the process and its ability to analyze large amounts of
likelihood of hallucinations. Retrieval-augmented retrieved data without losing important nuances.
generation works on the same principle, but instead Overall, LLM systems that include retrieval-
of the user inputting documents into the context augmented generation techniques are far more
window, the LLM does this work itself. When the reliable than LLMs on their own, but users should
user asks the LLM a question, the LLM will first still be aware that response outputs may still be
identify relevant terms in the user’s question and inaccurate or incomplete.
generate a search prompt. The LLM will then use
that prompt to query an Internet search engine One of the qualities that make LLMs compelling is
or specialized information database. Having their capability to engage in processes resembling
retrieved webpages or documents relevant to human creativity and problem solving, which
the user’s request, the LLM then incorporates would be hampered if models were strictly required
those documents into its context window, along to only produce completely accurate answers.
with the user’s original question/prompt, and As such, the possibility of hallucinations is not
generates a response taking all of that information
into account.
Retrieval-augmented generation has a number
of advantages, such as allowing the LLM
to access more up-to-date information or
specialized information databases and reducing
the likelihood of hallucinations, thus increasing
the factual reliability of LLM outputs. However,
retrieval-augmented generation is not a perfect
FOUNDATIONS OF AI | LOZANOSMITH.COM VOLUME 1 | 19
