作者：京東科技 賈世聞

RAG（Retrieval-Augmented Generation）技術在AI生態系統中扮演著至關重要的角色，特別是在提升大型語言模型（LLMs）的準確性和應用范圍方面。RAG通過結合檢索技術與LLM提示，從各種數據源檢索相關信息，并將其與用戶的問題結合，生成準確且豐富的回答。這一機制特別適用于需要應對信息不斷更新的場景，因為大語言模型所依賴的參數知識本質上是靜態的。

RAG技術的優勢在于它能夠利用外部知識庫，引用大量的信息，以提供更深入、準確且有價值的答案，提高了生成文本的可靠性。此外，RAG模型具備檢索庫的更新機制，可以實現知識的即時更新，無需重新訓練模型，這在及時性要求高的應用中占優勢。

目前構建一個RAG并不是一個非常的事情。使用Langchain等成熟技術架構百十行代碼就能構建一個Demo。那能不能利用目前的Rust生態構建一個簡易的RAG。說干就干，本期和大家聊聊如果使用rust語言構建rag。

構建知識庫

知識庫構建主要是模型+向量庫，為了保證所有系統中所有組件都使用rust構建，在限量數據庫的選型上我們使用qdrant,純rust構建的向量數據庫。

知識庫的構建最重要的步驟是embedding的過程。
過程如下：

模型加載

獲取文本token

通過模型獲取文本的Embedding
下面詳細介紹每個過程細節及代碼實現。

模型加載

以下代碼用于加載模型和tokenizer

async fn build_model_and_tokenizer(model_config: &ConfigModel) -> Result {
    let device = Device::new_cuda(0)?;
    let repo = Repo::with_revision(
        model_config.model_id.clone(),
        RepoType::Model,
        model_config.revision.clone(),
    );
    let (config_filename, tokenizer_filename, weights_filename) = {
        let api = ApiBuilder::new()    
            .build()?;
        let api = api.repo(repo);
        let config = api.get("config.json").await?;
        let tokenizer = api.get("tokenizer.json").await?;
        let weights = if model_config.use_pth {
            api.get("pytorch_model.bin").await?
        } else {
            api.get("model.safetensors").await?
        };
        (config, tokenizer, weights)A
    };
    let config = std::fs::read_to_string(config_filename)?;
    let mut config: Config = serde_json::from_str(&config)?;
    let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;

    let vb = if model_config.use_pth {
        VarBuilder::from_pth(&weights_filename, DTYPE, &device)?
    } else {
        unsafe { VarBuilder::from_mmaped_safetensors(&[weights_filename], DTYPE, &device)? }
    };
    if model_config.approximate_gelu {
        config.hidden_act = HiddenAct::GeluApproximate;
    }
    let model = BertModel::load(vb, &config)?;
    Ok((model, tokenizer))
}

模型和tokenizer是系統中頻繁調用的部分，所以為了避免重復加載，通過OnceCell構建靜態全局變量

pub static GLOBAL_EMBEDDING_MODEL: OnceCell> = OnceCell::const_new();

pub async fn init_model_and_tokenizer() -> Arc {
    let config = get_config().unwrap();
    let (m, t) = build_model_and_tokenizer(&config.model).await.unwrap();
    Arc::new((m, t))
}

在系統啟動時加載模型

GLOBAL_RUNTIME.block_on(async {
    log::info!("global runtime start!");
    // 加載model
    GLOBAL_EMBEDDING_MODEL
        .get_or_init(init_model_and_tokenizer)
        .await;
});

Embedding 過程主要由一下函數實現。

pub async fn embedding_setence(content: &str) -> Result>> {
    let m_t = GLOBAL_EMBEDDING_MODEL.get().unwrap();
    let tokens = m_t
        .1
        .encode(content, true)
        .map_err(E::msg)?
        .get_ids()
        .to_vec();
    let token_ids = Tensor::new(&tokens[..], &m_t.0.device)?.unsqueeze(0)?;
    let token_type_ids = token_ids.zeros_like()?;
    let sequence_output = m_t.0.forward(&token_ids, &token_type_ids)?;
    let (_n_sentence, n_tokens, _hidden_size) = sequence_output.dims3()?;
    let embeddings = (sequence_output.sum(1)? / (n_tokens as f64))?;
    let embeddings = normalize_l2(&embeddings)?;
    let encodings = embeddings.to_vec2::()?;
    Ok(encodings)
}

函數通過tokenizer encode輸入的文本，再使用模型embed token 獲取一個三維的Tensor，最后歸一化張量。

數據入庫

知識庫構建是將待檢索文本向量化后存儲到向量數據庫的過程。
本次使用京東云文檔作為原始文本，加工為以下格式。數據加工過程這里就不累述了。

{
    "content": "# 服務計費nn主機遷移服務自身為免費服務，但是遷移目標為云主機鏡像時，遷移過程依賴系統自動創建的 中轉資源的配合，這些資源中涉及部分付費資源，會產生相應費用。nn遷移過程涉及的中轉資付費資源配置及計費說明如下（單個遷移任務）：nn|  | 云主機 | 云硬盤 | 彈性公網IP |n| --- | --- | --- | ------ |n| 計費類型 | 按配置 | 按配置 | 按用量 |n| 規格配置 | 2C4G （c.n2.large或c.n3.large或c.n1.large） | 系統盤：40G 通用型SSD 數據盤：通用型SSD，數量及容量取決于源服務器系統盤及數據盤情況 | 30Mbps |n| 費用預估 | 云主機規格每小時價格\\*遷移時長 | 云硬盤規格每小時價格\\*遷移時長 | 彈性公網IP每小時保有費\\*遷移時長 僅使用彈性公網IP入方向流量，只涉及IP保有用，不涉及流量費用 |nn> 提示：n>n> * 遷移時長取決于源服務器遷數據量以及源服務器公網出方向帶寬，公網連接順暢且源服務器公網出方向帶寬不低于22.5Mbps的情況下（主機遷移為單線程傳輸，京東云云主機在單流傳輸下實際帶寬為帶寬上限的75%左右），實際數據容量為5GB的磁盤遷移時長在30分鐘左右。n> * 中轉實例實例綁定的安全組出方向默認拒絕所有流量，因此默認情況下降不會產生任何公網出方向收費流量，但此配置也影響了云主機部分監控指標的上報，如需要監控中轉實例的全部監控數據，可自行調整安全組規則方向出方向443端口。",
    "title": "服務計費說明",
    "product": "云主機 CVM",
    "url": "https://docs.jdcloud.com/cn/virtual-machines/server-migration-service/billing"
}

入庫完整代碼如下：

use anyhow::Error as E;
use anyhow::Result;
use candle_core::Device;
use candle_core::Tensor;
use candle_nn::VarBuilder;
use candle_transformers::models::bert::{BertModel, Config, HiddenAct, DTYPE};
use hf_hub::{api::tokio::Api, Repo, RepoType};
use qdrant_client::qdrant::CollectionExistsRequest;
use qdrant_client::qdrant::CreateCollectionBuilder;
use qdrant_client::qdrant::DeleteCollection;
use qdrant_client::qdrant::Distance;
use qdrant_client::qdrant::UpsertPointsBuilder;
use qdrant_client::qdrant::VectorParamsBuilder;
use qdrant_client::Payload;
use qdrant_client::{
    qdrant::{
        CollectionOperationResponse, CreateCollection, PointStruct, PointsOperationResponse,
        UpsertPoints,
    },
    Qdrant,
};
use serde::{Deserialize, Serialize};
use serde_json::from_str;
use std::fs;
use std::sync::Arc;
use tokenizers::Tokenizer;
use tokio::sync::OnceCell;
use uuid::Uuid;
use walkdir::WalkDir;

#[derive(Debug, Serialize, Deserialize, Clone)]
pub struct Doc {
    pub content: String,
    pub title: String,
    pub product: String,
    pub url: String,
}

#[derive(Debug, PartialEq, Serialize, Deserialize, Clone)]
pub struct ModelConfig {
    #[serde(default = "ModelConfig::model_id_default")]
    pub model_id: String,
    #[serde(default = "ModelConfig::revision_default")]
    pub revision: String,
    #[serde(default = "ModelConfig::use_pth_default")]
    pub use_pth: bool,
    #[serde(default = "ModelConfig::approximate_gelu_default")]
    pub approximate_gelu: bool,
}

impl Default for ModelConfig {
    fn default() -> Self {
        Self {
            model_id: Self::model_id_default(),
            revision: Self::revision_default(),
            use_pth: Self::use_pth_default(),
            approximate_gelu: Self::approximate_gelu_default(),
        }
    }
}

impl ModelConfig {
    fn model_id_default() -> String {
        "moka-ai/m3e-large".to_string()
    }
    fn revision_default() -> String {
        "main".to_string()
    }
    fn use_pth_default() -> bool {
        true
    }
    fn approximate_gelu_default() -> bool {
        false
    }
}

pub static GLOBAL_MODEL: OnceCell> = OnceCell::const_new();
pub static GLOBAL_TOKEN: OnceCell> = OnceCell::const_new();

pub async fn init_model() -> Arc {
    let config = ModelConfig::default();
    let (m, _) = build_model_and_tokenizer(&config).await.unwrap();
    Arc::new(m)
}

pub async fn init_tokenizer() -> Arc {
    let config = ModelConfig::default();
    let (_, t) = build_model_and_tokenizer(&config).await.unwrap();
    Arc::new(t)
}

async fn build_model_and_tokenizer(model_config: &ModelConfig) -> Result {
    let device = Device::new_cuda(0)?;
    let repo = Repo::with_revision(
        model_config.model_id.clone(),
        RepoType::Model,
        model_config.revision.clone(),
    );
    let (config_filename, tokenizer_filename, weights_filename) = {
        let api = Api::new()?;
        let api = api.repo(repo);
        let config = api.get("config.json").await?;
        let tokenizer = api.get("tokenizer.json").await?;
        let weights = if model_config.use_pth {
            api.get("pytorch_model.bin").await?
        } else {
            api.get("model.safetensors").await?
        };
        (config, tokenizer, weights)
    };
    let config = std::fs::read_to_string(config_filename)?;
    let mut config: Config = serde_json::from_str(&config)?;
    let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;

    let vb = if model_config.use_pth {
        VarBuilder::from_pth(&weights_filename, DTYPE, &device)?
    } else {
        unsafe { VarBuilder::from_mmaped_safetensors(&[weights_filename], DTYPE, &device)? }
    };
    if model_config.approximate_gelu {
        config.hidden_act = HiddenAct::GeluApproximate;
    }
    let model = BertModel::load(vb, &config)?;
    Ok((model, tokenizer))
}

pub async fn embedding_setence(content: &str) -> Result>> {
    let m = GLOBAL_MODEL.get().unwrap();
    let t = GLOBAL_TOKEN.get().unwrap();
    let tokens = t.encode(content, true).map_err(E::msg)?.get_ids().to_vec();

    let token_ids = Tensor::new(&tokens[..], &m.device)?.unsqueeze(0)?;
    let token_type_ids = token_ids.zeros_like()?;

    let sequence_output = m.forward(&token_ids, &token_type_ids)?;
    let (_n_sentence, n_tokens, _hidden_size) = sequence_output.dims3()?;
    let embeddings = (sequence_output.sum(1).unwrap() / (n_tokens as f64))?;
    let embeddings = normalize_l2(&embeddings).unwrap();
    let encodings = embeddings.to_vec2::()?;
    Ok(encodings)
}

pub fn normalize_l2(v: &Tensor) -> Result {
    Ok(v.broadcast_div(&v.sqr()?.sum_keepdim(1)?.sqrt()?)?)
}

pub struct QdrantClient {
    client: Qdrant,
}

impl QdrantClient {
    pub async fn create_collection(
        &self,
        request: impl Into,
    ) -> Result {
        let resp = self.client.create_collection(request).await?;
        Ok(resp)
    }

    pub async fn delete_collection(
        &self,
        request: impl Into,
    ) -> Result {
        let resp = self.client.delete_collection(request).await?;
        Ok(resp)
    }

    pub async fn collection_exists(
        &self,
        request: impl Into,
    ) -> Result {
        let resp = self.client.collection_exists(request).await?;
        Ok(resp)
    }

    pub async fn load_dir(&self, path: &str, collection_name: &str) {
        let mut points = vec![];
        for entry in WalkDir::new(path)
            .into_iter()
            .filter_map(Result::ok)
            .filter(|e| !e.file_type().is_dir() && e.file_name().to_str().is_some())
        {
            if let Some(p) = entry.path().to_str() {
                let id = Uuid::new_v4();
                let content = match fs::read_to_string(p) {
                    Ok(c) => c,
                    Err(_) => continue,
                };

                let doc = match from_str::(content.as_str()) {
                    Ok(d) => d,
                    Err(_) => continue,
                };
                let mut payload = Payload::new();
                payload.insert("content", doc.content);
                payload.insert("title", doc.title);
                payload.insert("product", doc.product);
                payload.insert("url", doc.url);
                let vector_contens = embedding_setence(content.as_str()).await.unwrap();
                let ps = PointStruct::new(id.to_string(), vector_contens[0].clone(), payload);
                points.push(ps);

                if points.len().eq(&100) {
                    let p = points.clone();
                    self.client
                        .upsert_points(UpsertPointsBuilder::new(collection_name, p).wait(true))
                        .await
                        .unwrap();
                    points.clear();
                    println!("batch finish");
                }
            }
        }

        if points.len().gt(&0) {
            self.client
                .upsert_points(UpsertPointsBuilder::new(collection_name, points).wait(true))
                .await
                .unwrap();
        }
    }
}

#[tokio::main]
async fn main() {
    // 加載模型
    GLOBAL_MODEL.get_or_init(init_model).await;
    GLOBAL_TOKEN.get_or_init(init_tokenizer).await;

    let collection_name = "default_collection";

    // The Rust client uses Qdrant's GRPC interface
    let qdrant = Qdrant::from_url("http://localhost:6334").build().unwrap();
    let qdrant_client = QdrantClient { client: qdrant };

    if !qdrant_client
        .collection_exists(collection_name)
        .await
        .unwrap()
    {
        qdrant_client
            .create_collection(
                CreateCollectionBuilder::new(collection_name)
                    .vectors_config(VectorParamsBuilder::new(1024, Distance::Dot)),
            )
            .await
            .unwrap();
    }

    qdrant_client
        .load_dir("/root/jd_docs", collection_name)
        .await;

    println!("{:?}", qdrant_client.client.health_check().await);
}

以上代碼要完成的任務如下：

推理服務

推理服務使用 rust 構建的 mistral.rs。

由于國內訪問hf 并不方便所以先通過 https://hf-mirror.com/ 現將模型下載到本地。本次使用qwen模型

HF_ENDPOINT="https://hf-mirror.com"  huggingface-cli download --repo-type model --resume-download Qwen/Qwen2-7B --local-dir /root/Qwen2-7B

啟動 mistralrs-server

git clone https://github.com/EricLBuehler/mistral.rs
cd mistral.rs
cargo run  --bin mistralrs-server  --features cuda -- --port 3333 plain -m /root/Qwen2-7B  -a qwen2

推理服務調用

mistral.rs 支持 Openai 的 api接口，使用 openai-api-rs調用即可。推理時間比較長 timeout 要設置長一些，若timeout 時間太短有可能不等返回結果就已經強制超時。

pub static GLOBAL_OPENAI_CLIENT: Lazy> = Lazy::new(|| {
    let mut client =
        OpenAIClient::new_with_endpoint("http://10.0.0.7:3333/v1".to_string(), "EMPTY".to_string());
    client.timeout = Some(30);
    Arc::new(client)
});

pub async fn inference(content: &str, max_len: i64) -> Result> {
    let req = ChatCompletionRequest::new(
        "".to_string(),
        vec![chat_completion::ChatCompletionMessage {
            role: chat_completion::MessageRole::user,
            content: chat_completion::Content::Text(content.to_string()),
            name: None,
            tool_calls: None,
            tool_call_id: None,
        }],
    )
    .max_tokens(max_len);

    let cr = GLOBAL_OPENAI_CLIENT.chat_completion(req).await?;
    Ok(cr.choices[0].message.content.clone())
}

將Retriever和推理服務集成

pub async fn answer(question: &str, max_len: i64) -> Result> {
    let retriver = retriever(question, 1).await?;
    let mut context = "".to_string();

    for sp in retriver.result {
        let payload = sp.payload;
        let product = payload.get("product").unwrap().to_string();
        let title = payload.get("title").unwrap().to_string();
        let content = payload.get("content").unwrap().to_string();
        context.push_str(&product);
        context.push_str(&title);
        context.push_str(&content);
    }

    let prompt = format!(
        "你是一個云技術專家, 使用以下檢索到的Context回答問題。用中文回答問題。
        Question: {}
        Context: {}
        ",
        question, context
    );

    log::info!("{}", prompt);

    let req = ChatCompletionRequest::new(
        "".to_string(),
        vec![chat_completion::ChatCompletionMessage {
            role: chat_completion::MessageRole::user,
            content: chat_completion::Content::Text(prompt),
            name: None,
            tool_calls: None,
            tool_call_id: None,
        }],
    )
    .max_tokens(max_len);

    let cr = GLOBAL_OPENAI_CLIENT.chat_completion(req).await?;
    Ok(cr.choices[0].message.content.clone())
}

后記

完整工程地址[embedding_server]https://github.com/jiashiwen/embedding_server

后續工程問題，多卡推理，多機推理，推理加速

資源對比

GPU 型號

|=========================================+========================+======================|
|   0  NVIDIA A30                     Off |   00000000:00:07.0 Off |                    0 |
| N/A   30C    P0             29W /  165W |       0MiB /  24576MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+

Embedding 資源

m3e-large

vllm

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A    822789      C   ...iprojects/rag_demo/.venv/bin/python       1550MiB |
+-----------------------------------------------------------------------------------------+

candle

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A    823261      C   target/debug/embedding_server                1484MiB |
+-----------------------------------------------------------------------------------------+

推理資源

Qwen1.5-1.8B-Chat

vllm

|=========================================================================================|
|    0   N/A  N/A    822437      C   /usr/bin/python3                            20440MiB |
+-----------------------------------------------------------------------------------------+

mistral.rs

|=========================================================================================|
|    0   N/A  N/A    822174      C   target/debug/mistralrs-server               22134MiB |
+-----------------------------------------------------------------------------------------+

Qwen2-7B

vllm 現存溢出

[rank0]: OutOfMemoryError: CUDA out of memory. Tried to allocate 9.25 GiB. GPU

mistral.rs

|=========================================================================================|
|    0   N/A  N/A    656923      C   target/debug/mistralrs-server               22006MiB |
+-----------------------------------------------------------------------------------------+

從實際情況來看，Embedding 模型再資源占用情況 rust candle框架使用顯存略小些；推理模型Qwen1.5-1.8B-Chat，vllm 資源占用略小。Qwen2-7B vllm直接顯存溢出。

坑

大部分框架中使用 hf-hub 采用同步調用，不支持境內的mirror。動手改造

src/api/tokio.rs

impl ApiBuilder {
    /// Set endpoint example 'https://hf-mirror.com'
    pub fn with_endpoint(mut self, endpoint: &str) -> Self {
        self.endpoint = endpoint.to_string();
        self
    }
}

審核編輯 黃宇