multilingual-e5-small + ONNX Runtime，使用 ONNX Runtime 的内置分词

using System;

using System.Collections.Generic;

using System.IO;

using System.Linq;

using Microsoft.ML.OnnxRuntime;

using Microsoft.ML.OnnxRuntime.Tensors;

public class MultilingualE5Embedding : IDisposable

{

    private readonly InferenceSession _session;

    private const int EmbeddingSize = 384;

    private const int MaxSeqLen = 128;

    // ======================

    // 删掉所有单例代码！！！

    // ======================

    public MultilingualE5Embedding()

    {

        string modelPath = Path.Combine(

            AppContext.BaseDirectory,

            "MLModels",

            "multilingual-e5-small",

            "model.onnx"

        );

        var options = new SessionOptions();

        options.AppendExecutionProvider_CPU(0);

        options.IntraOpNumThreads = 1;

        _session = new InferenceSession(modelPath, options);

    }

    public float[] Encode(string text)

    {

        if (string.IsNullOrWhiteSpace(text))

            return new float[EmbeddingSize];

        string input = $"query: {text.Trim()}";

        long[] inputIds;

        long[] attentionMask;

        long[] tokenTypeIds;

        EncodeInput(input, out inputIds, out attentionMask, out tokenTypeIds);

        var inputs = new List<NamedOnnxValue>

        {

            NamedOnnxValue.CreateFromTensor("input_ids", ToTensor(inputIds)),

            NamedOnnxValue.CreateFromTensor("attention_mask", ToTensor(attentionMask)),

            NamedOnnxValue.CreateFromTensor("token_type_ids", ToTensor(tokenTypeIds))

        };

        using var results = _session.Run(inputs);

        var lastHidden = results[0].AsTensor<float>();

        float[] embedding = new float[EmbeddingSize];

        for (int i = 0; i < EmbeddingSize; i++)

            embedding[i] = lastHidden[0, 0, i];

        return embedding;

    }

    private void EncodeInput(string text, out long[] inputIds, out long[] attentionMask, out long[] tokenTypeIds)

    {

        List<long> ids = new List<long> { 0 };

        foreach (char c in text)

            ids.Add(CharToId(c));

        ids.Add(2);

        while (ids.Count < MaxSeqLen)

            ids.Add(1);

        if (ids.Count > MaxSeqLen)

            ids = ids.Take(MaxSeqLen).ToList();

        inputIds = ids.ToArray();

        attentionMask = inputIds.Select(i => i != 1 ? 1L : 0L).ToArray();

        tokenTypeIds = new long[MaxSeqLen];

    }

    private long CharToId(char c)

    {

        if (c >= 32 && c <= 126) return c + 100;

        return 3;

    }

    private DenseTensor<long> ToTensor(long[] array)

    {

        return new DenseTensor<long>(array, new[] { 1, array.Length });

    }

    public void Dispose() => _session?.Dispose();

}

// 向量工具类

public static class VectorHelper

{

    public static byte[] Serialize(float[] vector)

    {

        byte[] bytes = new byte[vector.Length * 4];

        Buffer.BlockCopy(vector, 0, bytes, 0, bytes.Length);

        return bytes;

    }

    public static float[] Deserialize(byte[] bytes)

    {

        float[] vector = new float[bytes.Length / 4];

        Buffer.BlockCopy(bytes, 0, vector, 0, bytes.Length);

        return vector;

    }

    public static float CosineSimilarity(float[] a, float[] b)

    {

        if (a == null || b == null || a.Length != b.Length) return 0;

        float dot = 0, magA = 0, magB = 0;

        for (int i = 0; i < a.Length; i++)

        {

            dot += a[i] * b[i];

            magA += a[i] * a[i];

            magB += b[i] * b[i];

        }

        return dot / (float)Math.Sqrt(magA * magB);

    }

}

using System;

using System.Collections.Generic;

using System.Linq;

using System.Text;

using System.Threading.Tasks;

using Microsoft.ML.Tokenizers;

using System.IO;

using System;

using System.Collections.Generic;

namespace 智能客服.Services

{

    public static class VectorHelper

    {

        // float[] → byte[] 存入数据库

        public static byte[] Serialize(float[] vector)

        {

            byte[] bytes = new byte[vector.Length * 4];

            Buffer.BlockCopy(vector, 0, bytes, 0, bytes.Length);

            return bytes;

        }

        // byte[] → float[] 从数据库读取

        public static float[] Deserialize(byte[] bytes)

        {

            float[] vector = new float[bytes.Length / 4];

            Buffer.BlockCopy(bytes, 0, vector, 0, bytes.Length);

            return vector;

        }

        // 余弦相似度

        public static float CosineSimilarity(float[] a, float[] b)

        {

            if (a == null || b == null || a.Length != b.Length) return 0;

            float dot = 0, magA = 0, magB = 0;

            for (int i = 0; i < a.Length; i++)

            {

                dot += a[i] * b[i];

                magA += a[i] * a[i];

                magB += b[i] * b[i];

            }

            return dot / ((float)Math.Sqrt(magA) * (float)Math.Sqrt(magB));

        }

    }

}

使用示例：

// 为问题生成向量（使用 "passage:" 前缀）

using var embed = new MultilingualE5Embedding();

float[] vec = embed.Encode(question); // 自动带 query: 前缀

byte[] blob = VectorHelper.Serialize(vec);