Merge pull request #1630 from MicrosoftDocs/main

PhilKang0704 · web-flow · commit 5f42b5a670f9 · 2024-11-20T13:30:33.000+08:00
11/20 11:00 AM IST Publish
diff --git a/articles/search/semantic-how-to-query-rewrite.md b/articles/search/semantic-how-to-query-rewrite.md
@@ -75,7 +75,7 @@ In this REST API example, we use [Search Documents](/rest/api/searchservice/docu
     - We set "semanticConfiguration" to a [predefined semantic configuration](semantic-how-to-configure.md) embedded in your index.
     - We set "queryType" to "semantic". We either need to set "queryType" to "semantic" or include a nonempty "semanticQuery" property in the request. [Semantic ranking](semantic-search-overview.md) is required for query rewriting.
     - We set "queryRewrites" to "generative|count-5" to get up to five query rewrites. You can set the count to any value between 1 and 10. 
-    - We set "queryLanguage" to the target language ("en-US") of the query rewrites. The supported locales are: 
+    - Since we requested query rewrites by setting the "queryRewrites" property, we must set "queryLanguage" to the search text language. The Search service uses the same language for the query rewrites. In this example, we use "en-US". The supported locales are: 
         `en-AU`, `en-CA`, `en-GB`, `en-IN`, `en-US`, `ar-EG`, `ar-JO`, `ar-KW`, `ar-MA`, `ar-SA`, `bg-BG`, `bn-IN`, `ca-ES`, `cs-CZ`, `da-DK`, `de-DE`, `el-GR`, `es-ES`, `es-MX`, `et-EE`, `eu-ES`, `fa-AE`, `fi-FI`, `fr-CA`, `fr-FR`, `ga-IE`, `gl-ES`, `gu-IN`, `he-IL`, `hi-IN`, `hr-BA`, `hr-HR`, `hu-HU`, `hy-AM`, `id-ID`, `is-IS`, `it-IT`, `ja-JP`, `kn-IN`, `ko-KR`, `lt-LT`, `lv-LV`, `ml-IN`, `mr-IN`, `ms-BN`, `ms-MY`, `nb-NO`, `nl-BE`, `nl-NL`, `no-NO`, `pa-IN`, `pl-PL`, `pt-BR`, `pt-PT`, `ro-RO`, `ru-RU`, `sk-SK`, `sl-SL`, `sr-BA`, `sr-ME`, `sr-RS`, `sv-SE`, `ta-IN`, `te-IN`, `th-TH`, `tr-TR`, `uk-UA`, `ur-PK`, `vi-VN`, `zh-CN`, `zh-TW`.
     - We set "debug" to "queryRewrites" to get the query rewrites in the response. 
   
diff --git a/articles/search/vector-search-how-to-storage-options.md b/articles/search/vector-search-how-to-storage-options.md
@@ -20,21 +20,21 @@ Azure AI Search stores multiple copies of vector fields that are used in specifi
 
 ## How vector fields are stored
 
-For every vector field, there are three copies of the vectors:
+For every vector field, there could be three copies of the vectors, each serving a different purpose:
 
-| Instance | Usage |
-|----------|-------|
-| Source vectors (in JSON) as received from an embedding model or push request to the index | Used for incremental data refresh, and if you want "retrievable" vectors returned in the query response. |
-| Original full-precision vectors | Unavailable or unsupported if vectors are uncompressed. Otherwise it's used for optional rescoring if query results obtained over compressed vectors. Rescoring applies only if vector fields undergo [scalar or binary quantization](vector-search-how-to-quantization.md). |
-| Vectors in the [HNSW graph for Approximate Nearest Neighbors (ANN) search](vector-search-overview.md) | Used for query execution. |
+| Instance | Usage | Controlled using |
+|----------|-------|------------|
+| Source vectors which store the JSON that was received during document indexing | Used for incremental data refresh with `merge` or `mergeOrUpload` during document indexing. Also used if you want "retrievable" vectors returned in the query response. | `stored` property on vector fields |
+| Original full-precision vectors | In existing indexes, these are used for internal index operations and for exhaustive KNN search. For vectors using compression, it's also used for rescoring (if enabled) on an oversampled candidate set of results from ANN search on vector fields using [scalar or binary quantization](vector-search-how-to-quantization.md) compression. | `rescoringOptions.rescoreStorageMethod` property in `vectorSearch.compressions`. For *uncompressed* vector fields on indexes created with `2024-11-01-Preview` API versions and later, this will be omitted by default with no impact on search activities nor quality. |
+| Vectors in the [HNSW graph for Approximate Nearest Neighbors (ANN) search](vector-search-overview.md) | Used for ANN query execution. Consists of either full-precision vectors (when no compression is applied) or quantized vectors (when compression is applied) | Only applies to HNSW. These data structures are required for efficient ANN search. |
 
 You can set properties that permanently discard the first two instances from vector storage.
 
-The last instance (vectors and graph) is required for ANN vector query execution. Lossy compression techniques like [scalar or binary quantization](vector-search-how-to-quantization.md) are applied to this vector instance. If you want to offset lossy compression, you should keep the second instance for rescoring purposes.
+The last instance (vectors and graph) is required for ANN vector query execution. If any compression techniques such as [scalar or binary quantization](vector-search-how-to-quantization.md) are used, they would be applied to this set of data. If you want to offset lossy compression, you should keep the second instance for rescoring purposes to improve ANN search quality.
 
 ## Set the `stored` property
 
-The `stored` property is a boolean on a vector field definition that determines whether storage is allocated for retrievable vector field content (the source instance). The `stored` property is true by default. If you don't need raw vector content in a query response, you can save up to 50 percent storage per field by changing `stored` to false.
+The `stored` property is a boolean property on a vector field definition that determines whether storage is allocated for retrievable vector field content (the source instance). The `stored` property is true by default. If you don't need raw vector content in a query response, you can save up to 50 percent storage per field by changing `stored` to false.
 
 Considerations for setting `stored` to false:
 
@@ -43,7 +43,7 @@ Considerations for setting `stored` to false:
 - However, if your indexing strategy includes [partial document updates](search-howto-reindex.md#update-content), such as "merge" or "mergeOrUpload" on an existing document, setting `stored=false` prevents content updates to those fields during the merge. On each "merge" or "mergeOrUpload" operation to a search document, you must provide the vector fields in its entirety, along with the nonvector fields that you're updating, or the vector is dropped.
 
 > [!IMPORTANT]
-> Setting the `stored=false` attribution is irreversible. It's set during index creation on vector fields when physical data structures are created. If you want retrievable vector content later, you must drop and rebuild the index, or create and load a new field that has the new attribution.
+> Setting the `stored=false` attribution is irreversible. This property can only be set when you create the index and is only allowed on vector fields. Updating an existing index with new vector fields cannot set this property to `false`. If you want retrievable vector content later, you must drop and rebuild the index, or create and load a new field that has the new attribution.
 
 For new vector fields in a search index, set `stored` to false to permanently remove retrievable storage for the vector field. The following example shows a vector field definition with the `stored` property.
 
@@ -81,9 +81,16 @@ PUT https://[service-name].search.windows.net/indexes/demo-index?api-version=202
 
 [!INCLUDE [Feature preview](./includes/previews/preview-generic.md)]
 
-The `rescoreStorageMethod` property on a vector field definition determines whether storage is allocated for original full-precision vectors. The `rescoreStorageMethod` property is set to `preserveOriginals` by default. If you aren't using the [oversampling and rescoring mitigations](vector-search-how-to-quantization.md#add-compressions-to-a-search-index) provided for querying compressed vectors, you can save on vector storage by changing `rescoreStorageMethod` to `discardOriginals`.
+The `rescoreStorageMethod` property controls the storage of full-precision vectors when compression is used.
 
-If you intend to use scalar or binary quantization, we recommend retaining `rescoreStorageMethod` set to `preserveOriginals`.
+For *uncompressed* vector fields on indexes created with `2024-11-01-Preview` API versions and later, this will be omitted by default with no impact on search activities nor quality. For existing vector fields created prior to this API version, there is no in-place ability to remove this copy of data.
+
+On a vector compression, the `rescoreStorageMethod` property is set to `preserveOriginals` by default, which retains full-precision vectors for[oversampling and rescoring capabilities](vector-search-how-to-quantization.md#add-compressions-to-a-search-index) to reduce the effect of lossy compression on the HNSW graph. If you don't use these capabilities, you can reduce vector storage by setting `rescoreStorageMethod` to `discardOriginals`.
+
+> [!IMPORTANT]
+> Setting the `rescoreStorageMethod` property is irreversible and will have different levels of search quality loss depending on the compression method. This can be set on indexes created with `2024-11-01-Preview` or later, either during index creation or adding new vector fields.
+
+If you intend to use scalar or binary quantization, we recommend retaining `rescoreStorageMethod` set to `preserveOriginals` to maximize search quality.
 
 To set this property:
 
