C#使用词嵌入向量与向量数据库为大语言模型(LLM)赋能长期记忆实现私域问答机器人落地

原始语料：ChatGLM是一个开源的清华技术成果转化的公司智谱AI研发的支持中英双语的对话机器人它支持中英双语问答的对话语言模型，基于 General Language Model (GLM 架构，具有 62 亿参数。结合模型量化技术，用户可以在消费级的显卡上进行本地部署（INT4 量化级别下最低只需 6GB 显存）。ChatGLM-6B 使用了和 ChatGLM 相同的技术，针对中文问答和对话进行了优化。经过约 1T 标识符的中英双语训练，辅以监督微调、反馈自助、人类反馈强化学习等技术的加持，62 亿参数的 ChatGLM-6B 已经能生成相当符合人类偏好的回答。

var embeddings = await new OpenAiOptions( { ApiKey = key }.Embeddings.CreateEmbedding(new EmbeddingCreateRequest( { InputAsList = inputs.ToList(, Model = OpenAI.GPT3.ObjectModels.Models.TextEmbeddingAdaV2 }; return embeddings.Data.Select(x => x.Embedding.ToList(;

[-0.0020597207,-0.012355088,0.0037828966,-0.032127112,-0.04815184,0.016633095,-0.01277577,........]

    public class ChatGlmVector
    {
        public ChatGlmVector(
        {
            Id = Id ?? Guid.NewGuid(.ToString(;
        }
        [Keyword]
        public string Id { get; set; }

        [Text]
        public string Text { get; set; }

        [DenseVector(Dimensions = 1536]
        public IList<double> Vector { get; set; }
    }

if (!elasticClient.Indices.Exists(IndexName.Exists elasticClient.Indices.Create(IndexName, c => c.Map<ChatGlmVector>(m => m.AutoMap(; await elasticClient.IndexAsync(item, idx => idx.Index(IndexName;

POST /my_index/_search
{
  "size": 3,    // 返回前3个最相似的文档
  "query": {
    "function_score": {
      "query": {
        "match_all": {}
      },
      "functions": [
        {
          "script_score": {
            "script": {
              "source": "def cosineSim = cosineSimilarity(params.queryVector, 'vector'; if (cosineSim > 0.8 return cosineSim; else return 0;",
              "params": {
                "queryVector": [1.0, 2.0, 3.0]   // 要查询的向量
              }
            }
          }
        }
      ],
      "boost_mode": "replace"
    }
  }
}

var scriptParams = new Dictionary<string, object> { {"queryVector", new double[]{1.0, 2.0, 3.0}} }; var script = new InlineScript("def cosineSim = cosineSimilarity(params.queryVector, 'vector'; if (cosineSim > 0.8 return cosineSim; else return 0;" { Params = scriptParams }; var searchResponse = client.Search<object>(s => s .Size(3 .Query(q => q .FunctionScore(fs => fs .Query(qq => qq .MatchAll( .Functions(fu => fu .ScriptScore(ss => ss .Script(sc => script .BoostMode(FunctionBoostMode.Replace ;

            return (await GetOpenAIService(.ChatCompletion.CreateCompletion(new ChatCompletionCreateRequest(
            {
                Messages=new List<ChatMessage>( {
                ChatMessage.FromUser("你是一个智能助手，你需要根据下面的事实依据回答问题。如果用户输入不在事实依据范围内，请说\"抱歉,这个问题我不知道。\"",
ChatMessage.FromUser($"事实依据：{这里需要从ES查询出相似度最高的文本作为LLM的长期记忆}",
ChatMessage.FromUser($"用户输入：{这里是用户的原始问题}"
                },
                Model = OpenAI.GPT3.ObjectModels.Models.ChatGpt3_5Turbo
            }.Choices.FirstOrDefault(.Message;