Relationship Between Search Results (value) and Semantic Answers (@search.answers) in Azure AI Search's Vector + Semantic Search

Question

Subject: Relationship Between Search Results (value) and Semantic Answers (@search.answers) in Azure AI Search's Vector + Semantic Search

I would like to ask a question regarding the specifications of the semantic answers output when performing "vector + semantic search" using Azure AI Search, as there are parts I do not understand.

I understand that "vector + semantic search" adjusts the scores of multiple data entries searched through vector search based on the "semantic relevance of the search query and data" using AI, calculates the "rerankerScore," and returns the results sorted by "rerankerScore."

Also, I understood that the semantic answer (@search.answers) returns the passage best suited as an answer to the query from the search results above.

However, upon executing the following search query, I found that the semantic answer is derived from a document not present in the search results (value). Does this mean that the "semantic answer" extracts the optimal answer from the entire document stored as an index, irrespective of the "score" or "rerankerScore," and not just from the "search results (value)"?

In that case, is it possible to extract the "semantic answer" specifically from the "search results (value)"?

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

{
"search": "介護休暇はだれが取得できる？",
"count": true,
"vectorQueries": [
{
"kind": "text",
"text": "介護休暇はだれが取得できる？",
"fields": "titleVector"
}
],
"queryType": "semantic",
"semanticConfiguration": "default",
"captions": "extractive",
"answers": "extractive|count-3",
"queryLanguage": "ja-jp",
"select": "chunk_id,parent_id,content,title,url,filepath",
"top": 5
}

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Answer 1

I apologize for my delayed response.

Thank you for your response.

I greatly appreciate the detailed explanation regarding the semantic answers in Azure AI Search. Your explanation clarified that the candidate pool referred to by the semantic ranker comprises the "top 50 results from the hybrid search" and not the "outputted results in value[]." This provides a clearer understanding of why the semantic answer may derive from documents not included in the Value results.

Additionally, regarding "setting a lower top value," does this refer to controlling the number of results passed to the reranker candidate pool, rather than the top value of the search query? If there are any resources or documentation available on this topic, I would appreciate it if you could share them with me.

Furthermore, concerning the method "pre-filtering the set of documents passed to semantic re-ranking," does this relate to the Prefiltering and Postfiltering mentioned in the URL below?

https://learn.microsoft.com/en-us/azure/search/hybrid-search-how-to-query#semantic-hybrid-search-with-filter

I apologize for asking additional questions while you are busy, but I would be grateful for your continued assistance. Thank you in advance for your help.