From bb81651ec013b7d2654f1cac8ebd38c5d5c92eae Mon Sep 17 00:00:00 2001
From: Wenbo Cao <104199@smsassist.com>
Date: Fri, 11 Aug 2023 14:16:08 -0500
Subject: [PATCH] modify cosine function

---
 .../Plugins/MemVecDb/MemVectorDatabase.cs     | 60 ++++++++++++++++---
 1 file changed, 53 insertions(+), 7 deletions(-)

diff --git a/src/Infrastructure/BotSharp.Core/Plugins/MemVecDb/MemVectorDatabase.cs b/src/Infrastructure/BotSharp.Core/Plugins/MemVecDb/MemVectorDatabase.cs
index e724e2eec..5aafbacf4 100644
--- a/src/Infrastructure/BotSharp.Core/Plugins/MemVecDb/MemVectorDatabase.cs
+++ b/src/Infrastructure/BotSharp.Core/Plugins/MemVecDb/MemVectorDatabase.cs
@@ -1,5 +1,7 @@
 using BotSharp.Abstraction.VectorStorage;
+using Tensorflow;
 using Tensorflow.NumPy;
+using static Tensorflow.Binding;
 
 namespace BotSharp.Core.Plugins.MemVecDb;
 
@@ -60,18 +62,62 @@ private float[] CalEuclideanDistance(float[] vec, List<VecRecord> records)
         return c.ToArray<float>();
     }
 
-    public float[] CalCosineSimilarity(float[] vec, List<VecRecord> records)
+    public NDArray CalCosineSimilarity(float[] vec, List<VecRecord> records)
     {
-        var similarities = new float[records.Count];
-        var a = vec;
-        var normA = np.linalg.norm(a);
+        var recordsArray = np.zeros((records.Count, records[0].Vector.Length), dtype: np.float32);
 
         for (int i = 0; i < records.Count; i++)
         {
-            var b = records[i].Vector;
-            similarities[i] = np.dot(a, b) / (normA * np.linalg.norm(b));
+            recordsArray[i] = records[i].Vector;
         }
 
-        return similarities;
+        var vecArray = np.expand_dims(np.array(vec, dtype: np.float32), axis: 0); // [1. 300]
+
+        (var normVecArray, var _) = SafeNormalize(vecArray);
+        (var normRecordsArray, var _) = SafeNormalize(recordsArray);
+
+        var simiMatix = tf.matmul(tf.cast(normVecArray, tf.float32), tf.transpose(tf.cast(normRecordsArray, tf.float32))).numpy(); // [1, num_records]
+
+        simiMatix = np.squeeze(simiMatix, axis: 0);
+
+        return simiMatix;
+    }
+
+    public int[] CalCosineSimilarityTopK(float[] vec, List<VecRecord> records, int topK = 10, float filterProb = 0.75f)
+    {
+        var simiMatix = CalCosineSimilarity(vec, records);
+
+        var topIndex = np.argsort(simiMatix)["::-1"][$":{topK}"];
+
+        var resIndex = new List<int>();
+
+        for (int i = 0; i < topK; i++)
+        {
+            var index = topIndex[i];
+            var value = simiMatix[index];
+
+            if (value > filterProb)
+            {
+                resIndex.Add(topIndex[i]);
+            }
+        }
+
+        return resIndex.ToArray();
+    }
+
+    public (NDArray, NDArray) SafeNormalize(NDArray x, double eps = 2.223E-15)
+    {
+        var squaredX = np.sum(np.multiply(x, x), axis: 1);
+        var normX = np.sqrt(squaredX);
+
+        var epsTensor = tf.cast(tf.convert_to_tensor(eps), dtype: tf.float32);
+        var normXTensor = tf.cast(normX, tf.float32);
+        var contantMask = (normXTensor < epsTensor);
+        var divideTensor = tf.ones_like(normXTensor, dtype: tf.float32);
+
+        normX = tf.where(contantMask, divideTensor, normXTensor).numpy();
+        normX = np.expand_dims(normX, axis: 1);
+
+        return (x / normX, normX);
     }
 }