update READMEs

Author: masajiro

README-jp.md

@@ -10,6 +10,8 @@ Neighborhood Graph and Tree for Indexing High-dimensional Data
 
 ニュース
 -------
+- 05/26/2025 直積量子化がグラフのデータタイプとして利用可能になりました。 (v2.4.0)
+- 10/29/2024 スカラ量子化がグラフやQBGのデータタイプとして利用可能になりました。(v2.3.0)
 - 2024/04/10 内積が利用可能になりました。(v2.1.0)
 - 2022/08/10 QBG(Quantized Blob Graph)およびQG(NGTQGの改良版)が利用可能となりました。ngtqおよびngtqgは[qbg](https://github.com/yahoojapan/NGT/blob/main/bin/qbg/README.md)で置き換えられました。
 - 2022/02/04 FP16(半精度浮動小数点)が利用可能になりました。(v1.14.0)

README.md

@@ -10,6 +10,8 @@ Neighborhood Graph and Tree for Indexing High-dimensional Data
 
 News
 ----
+- 05/26/2025 Product quantization is now available as a data type for graph indices. (v2.4.0)
+- 10/29/2024 Scalar quantization is now available as a data type for graph indices and QBG. (v2.3.0)
 - 04/10/2024 Inner product (or dot product) is now available. (v2.1.0)
 - 08/10/2022 [QBG](https://github.com/yahoojapan/NGT#qbg-quantized-blob-graph-based-method) (Quantized Blob Graph) and [QG](https://github.com/yahoojapan/NGT#qg-quantized-graph-based-method) (renewed NGTQG) are now available. The command-line interface ngtq and ngtqg are now obsolete by replacing [qbg](bin/qbg/README.md). (v2.0.0)
 - 02/04/2022 FP16 (half-precision floating point) is now available. (v1.14.0)

bin/ngt/README-jp.md

@@ -32,9 +32,10 @@ CygWin といった POSIXLY_CORRECT が設定されている環境では、コ
 
 -   *[create](#create)*
 -   *[append](#append)*
+-   *[prep-pq](#prep-pq)*
 -   *[search](#search)*
 -   *[remove](#remove)*
--   *[prune](#prune)*
+-   *[prune](#prune-非推奨)*
 -   *[reconstruct graph](#reconstruct-graph)*
 -   *[rebuild](#rebuild)*
 
@@ -93,6 +94,7 @@ ANNGやBKNNGを指定した場合には登録データ（ノード）からエ
 - __f__: 4バイト浮動小数点（デフォルト）
 - __h__: 2バイト浮動小数点
 - __q__: 1バイトスカラー量子化
+- __pq4__: 4ビット直積量子化
 
 **-D** *distance\_function*  
 距離関数を指定します。
@@ -134,6 +136,17 @@ ANNGやBKNNGを指定した場合には登録データ（ノード）からエ
 **-n** *no\_of\_registration\_data*  
 登録するデータ数を指定します。指定しない場合には指定されたファイル中のすべてのデータを登録します。
 
+### PREP-PQ
+
+プロダクト量子化（product quantization）データ型の前処理を実行します。
+
+      $ ngt prep-pq index
+
+
+*index*  
+データ型がプロダクト量子化に設定され、オブジェクトが精査オブジェクトリポジトリに追加されている既存のインデックス名を指定してください。
+このコマンドは、オブジェクトを量子化するためのQBGデータ構造を構築します。
+
 ### SEARCH
 
 指定されたクエリデータを用いてインデックスを検索します。

bin/ngt/README.md

@@ -34,6 +34,7 @@ Description
 
 -   *[create](#create)*
 -   *[append](#append)*
+-   *[prep-pq](#prep-pq)*
 -   *[search](#search)*
 -   *[remove](#remove)*
 -   *[prune](#prune-not-recommended)*
@@ -95,6 +96,7 @@ Specify the data object type.
 - __f__: 4 byte floating point number (default)
 - __h__: 2 byte floating point number
 - __q__: 1 byte scalar quantization
+- __pq4__: 4 bit product quantization
 
 **-D** *distance\_function*  
 Specify the distance function as follows.
@@ -137,6 +139,17 @@ Specify the number of dimensions of registration data. Specification is unnecess
 **-n** *no\_of\_registration\_data*  
 Specify the number of data items to be registered. If not specified, all data in the specified file will be registered.
 
+### PREP-PQ
+
+Execute preprocessing for the product quantization data type.
+
+      $ ngt prep-pq index
+        
+
+*index*  
+Specify the name of the existing index where the data type is set to product quantization and the objects are appended to the refinement object repository. This command builds a QBG data structure to quantize the objects.
+
+
 ### SEARCH
 
 Search the index using the specified query data.

bin/qbg/README.md

@@ -158,7 +158,7 @@ Specify the number of trials to optimize the subvector quantization.
 
 Search the index using the specified query data.
 
-      $ qbg search [-n number_of_search_objects] [-e search_range_coefficient] [-p result_expansion]
+      $ qbg search [-n number_of_search_objects] [-e search_range_coefficient] [-p result_expansion] [-R refinement_data_type]
           index query_data
         
 
@@ -181,8 +181,12 @@ Specify the magnification coefficient (epsilon) of the search range for the quan
 Specify the number of the explored nodes in the graph. When the number of the explored nodes reached the specified number, the search is terminated. 
 
 **-p** *result_expansion* (default = 0.0)   
-Specify the expansion ratio of the number of approximate inner search objects to the number of search objects. For example, when the ratio is 10 and the number of search objects is 20, the number of the approximate search objects is set to 200 inside the search processing. A larger value brings higher accuracy but slower searching.
+Specify the expansion ratio of the number of approximate inner search objects to the number of search objects. For example, when the ratio is 10 and the number of search objects is 20, the number of approximate search objects is set to 200 inside the search process. After retrieving the approximate search objects, exact distances are calculated, and the objects are ranked to obtain the top results according to the number of search objects. This process is called search refinement. A larger value brings higher accuracy but slower searching. If the -R option is not specified, the objects for refinement are taken from secondary storage, which may slow down the refinement process.
 
+**-R** *refinement_data_type*
+Specify the data type for the search refinement. The objects for refinement are loaded into memory as the specified data type from secondary storage to shorten the refinement time. On the other hand, a large amount of memory may be used.
+- __f__: 4 byte floating point number (default)
+- __h__: 2 byte floating point number
 
 Examples of using the quantized graph
 -------------------------------------