Inside language models · Module 1: Inside language models in practice
Embeddings: similar does not mean equivalent
Learning outcome: Identify when semantic similarity is insufficient.
An embedding represents an item as a vector: a list of numbers used for operations such as similarity search. Related phrases can be near one another in that representation. Similarity is useful for finding candidates, but it is not proof that two claims have the same meaning.
Negation, quantities and product identifiers are particularly useful stress tests. “Refunds are permitted” and “refunds are not permitted” share much of their language while requiring opposite actions. A retrieval system must check the selected passage and its scope after similarity search. Think of the embedding as a way to find possible evidence, followed by a separate decision about whether it supports the answer.
Worked example
A fictional helpdesk search returns “The adapter supports model AX-10” for a question about AX-11. The passage is close in topic but does not establish compatibility. The correct response identifies the mismatch and seeks the AX-11 specification. Rewording the answer more confidently would hide the unresolved issue.
Try it yourself
Create four pairs of near-matching sentences. Change a negation, a model number, a date and a unit. State how each change affects the action a reader should take.
You can do this exercise on paper. If using an assistant, keep to the fictional material and an approved tool.
Compare your approach
The test pairs should be linguistically similar but decision-relevant. Keep each original and changed sentence together so a reviewer can see exactly which distinction retrieval must preserve.
Check your understanding
A semantically similar passage has the wrong product code. Is it sufficient evidence?
- Yes, because similarity means equivalence
- Yes, if two assistants return it
- No; verify the exact product and claim
Show the answer and explanation
Answer: No; verify the exact product and claim. Similarity retrieves candidates; it does not verify compatibility. A passage about AX-10 cannot establish an AX-11 claim without evidence connecting the products.
Workbook
Download the course workbook to record your answer and compare it after the knowledge check.