LEVEL 01 · LESSON 05
Embedding models vs LLMs
One model can find the evidence while another turns it into an answer.
What it is
An embedding model produces representations useful for comparison. A generative large language model produces sequences, usually text. Both learn patterns, but their outputs and jobs differ.
Why it matters
A search-and-answer system often needs both. Choosing a strong writing model does not automatically improve the search that supplies its evidence.
A business example
An internal assistant first uses an embedding model to find relevant expense-policy passages. A generative model then reads those passages and explains the reimbursement rules in plain language.
How it actually works
The embedding model maps each document chunk to a vector when it is indexed and maps a new query when it is searched. The generative model receives the retrieved text, not necessarily the vectors themselves. LLMs also use internal representations, but those are not interchangeable with the search embeddings your database expects. Changing the embedding model commonly requires re-embedding the document collection and measuring retrieval again.
A common misconception
An embedding model does not contain a precomputed vector for every possible sentence. It learns a function that produces representations for new inputs.
ONE SENTENCE TO REMEMBER
One model can find the evidence while another turns it into an answer.
CHECK YOUR UNDERSTANDING
Your company switches its document search to a new embedding model. What typically needs to happen next?
Show the answer
Answer: D.
Documents and queries must be embedded in compatible spaces, so vectors from the old model cannot be mixed with the new one. An embedding model learns a function for new inputs rather than storing a vector for every sentence. The generative model receives retrieved text, not the vectors.
Take it into your next meeting
If answers are fluent but use the wrong documents, which part would you investigate first?
One more concept, understood.
Mark this lesson complete to keep your place.