Sources checked
Embed questions and passages with compatible settings
In an original article search example, a question and document passages are embedded using compatible settings. The system retrieves nearby passages, then a reader or another component checks whether they answer the question.
Changing the model can invalidate an index
Changing the embedding model can change the vector space. Do not mix new query vectors with an old document index without verifying compatibility or rebuilding the index.
What the vector represents
An embedding model maps a string to a vector of fixed length. Training pushes texts that people or data labeled as related toward each other, and pushes unrelated texts apart.
Cosine similarity, or a dot product, then ranks neighbors. The vector has no column you can read as “this word means that.”
The model, the pooling method, and the input length all change the vector. You cannot mix vectors from two embedding models in one index and expect the distances to mean anything.
A query vector has to come from the same model that embedded the documents, including the same normalization.
Where the shortcut fails
Neighbors can share a topic and still disagree. A vector for “the drug was withdrawn” can sit near “the drug was approved” because both passages are about the same drug.
A search that returns the neighbor has not checked the claim. A person, or a second model with the passage in front of it, has to read the words.
Short queries, titles, and full paragraphs do not always live on the same scale. A product name can match a navigation menu instead of the policy you wanted.
Test with the queries your users type, including typos and very short phrases. Report recall on a labeled set before you call the index semantic search in a product sense.
| You have | You can conclude | You cannot conclude |
|---|---|---|
| A high cosine score | The model placed these strings near each other | The strings agree, or one cites the other |
| A shared topic neighbor | The index is a useful shortlist | The top hit is the right passage |
| Vectors from two models | Each space has its own geometry | The numbers are comparable |
What to record before you rely on it
Name the embedding model, the vector size, and the similarity function in the index documentation. State that vectors from any other model are not comparable.
Keep a small labeled set of queries with the passage a person accepted, and record how often that passage appears in the top five. When you upgrade the embedding model, rebuild the index.
Do not append new vectors to an old space and call it an incremental update. Keep the model id, the date, and the inputs you used beside that note, so a later change can be compared with the same test.
Common questions
Can I compare a cosine of 0.8 across two vendors?
No. The score depends on the model and how the vectors were normalized. Compare ranks inside one index, or compare models on the same labeled questions.
Do embeddings replace a reranker or a reader?
They propose candidates cheaply. A reranker or a person still has to check the wording, especially when a negation or a date changes the meaning.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





