When people hear the term AI hallucination, they often imagine an LLM confidently inventing facts that do not exist.
For example:
"The capital of France is Berlin."
That's an obvious hallucination because the answer is factually incorrect.
However, in real-world AI systems, hallucinations are often much more subtle.
Recently, I experienced a perfect example during the conversation with an AI assistant.
The Question
I asked:
"Is Redis a vector database?"
The assistant immediately responded:
"Yes, Redis is a vector database."
At first glance, the answer seemed reasonable.
After all, Redis supports:
- Vector storage
- Vector indexing
- Similarity search
These are all capabilities associated with vector databases.
But that wasn't actually what I was asking.
The Hidden Problem
My real question was:
"How is Redis classified as a database technology?"
In database classification terms, Redis is primarily:
- An in-memory database
- A key-value database
- A multi-model database
It is not generally classified as a dedicated vector database.
The assistant answered a different question:
"Can Redis be used as a vector database?"
instead of:
"Is Redis fundamentally a vector database?"
The result was interesting.
The answer contained correct facts.
Yet the answer was still wrong for the user's intent.
Why This Happens
Large Language Models (LLM) do not truly understand questions the way humans do.
Instead, they predict the most probable interpretation based on patterns learned during training.
When the model saw:
"Is Redis the vector database?"
it likely mapped the question to a common pattern:
"Can Redis perform a vector database functions?"
Since the answer to that interpretation is yes, the model confidently responded with "Yes."
The failure wasn't in factual knowledge.
The failure was in understanding the user's intent.
This Is a Form of Hallucination
Many teams define hallucinations as:
"Any output that does not correctly satisfy the user's request."
Under this broader definition, the Redis example qualifies.
The model generated an answer that was:
- Factually supported
- Logically consistent
- Yet misaligned with the actual question
In other words:
The model hallucinated the meaning of the question.
Why RAG Doesn't Fully Solve This
Many people believe that Retrieval-Augmented Generation (RAG) eliminates hallucinations.
But consider this scenario.
Even if a RAG system retrieves perfect documentation about Redis:
- Redis is an in-memory database
- Redis supports vector search
- Redis supports KNN queries
The LLM still has to interpret the user's question.
If it misunderstands the intent, it may still generate the wrong answer despite having perfect information.
This highlights an important reality:
Not all hallucinations come from missing knowledge.
Some hallucinations come from incorrect interpretation.
The Key Takeaway
When evaluating AI systems, don't only ask:
"Did the model know the answer?"
Also ask:
"Did the model understand the question?"
Because sometimes the most dangerous hallucinations are not invented facts.
They are correct facts applied to the wrong interpretation.
And from a user's perspective, the result is still an incorrect answer.
Top comments (1)
Strong framing. The distinction between factual errors and interpretive errors is underrated, especially in production AI workflows.
What stood out to me is your point that RAG can still fail when retrieval is correct but the model applies the wrong semantic lens. That matches a lot of what we see in real AI products: the problem is often not missing context, but mis-weighted context.
I also like that this pushes the conversation beyond the simplistic "hallucination = made-up fact" definition.