ragflow之初体验

最新推荐文章于 2024-10-11 19:42:09 发布

老鸟学AI

最新推荐文章于 2024-10-11 19:42:09 发布

阅读量800

点赞数 15

文章标签：语言模型云计算

本文链接：https://blog.csdn.net/smartvap_2012/article/details/142858507

版权

随着国内大语言模型（LLM）的应用日益广泛，构建面向企业私有数据的模型已成为一个重要研究方向。基于检索增强生成（RAG）的知识库与大语言模型的混合编排，为这一过程提供了强大的加速器，使得私有化部署更加高效和便捷。

ragflow 是RAG领域中较为理想的开源解决方案，在企业私有数据集成管理和知识库构建方面展现出显著优势，其灵活的架构设计和高效的知识库构建能力，使得数据检索和信息生成更加精准和高效。

本文将分享ragflow的部署方法和使用心得，欢迎各位批评指正。

1 环境准备

类型	版本
GPU	Tesla V100S / 32G x 8
CPU	Intel(R) Xeon(R) Gold 5218 / 64c
内存	376G
硬盘	SATA 4T
OS	CentOS7
体系结构	x64
CUDA	11.4
公网访问	代理

由于ragflow依赖的docker版本至少为24.0.9，而现有的GPU主机上还运行了其他低版本的docker容器，为了不影响这些存量应用容器，我们考虑二进制堆叠搭建一个新的docker运行环境，目标运行环境各组件的版本如下：

组件名称	版本
runc	1.1.12
docker-proxy	24.0.9
docker-init	0.19.0
dockerd	24.0.9
docker	24.0.9
ctr	1.7.13
containerd-shim-runc-v2	1.7.13
containerd	1.7.13
docker-compose	2.26.1

【注意】docker运行环境、docker-compose的二进制版本从此处下载：https://download.docker.com/linux/static/stable/x86_64/、https://github.com/docker/compose/releases?expanded=true&page=2&q=

1.1 containerd环境构建

1.1.1 containerd配置文件

将docker二进制文件解压至/data/labs/runtimes/docker-24/目录下，生成一个默认配置文件，并在此基础上修改。

./containerd config default > containerd-config.toml

【注意】文件完整内容如下，所有路径应考虑显式设置，尽量避免使用默认值，以防与已有docker运行环境产生冲突！

>>> containerd-config.toml <<<

disabled_plugins = []

imports = []

oom_score = 0

plugin_dir = "/data/labs/runtimes/docker-24/containerd-plugin"

required_plugins = []

root = "/data/labs/runtimes/docker-24/containerd-root"

state = "/data/labs/runtimes/docker-24/containerd-state"

temp = "/data/labs/runtimes/docker-24/containerd-temp"

version = 2

[cgroup]

path = "/data/labs/runtimes/docker-24/sys-fs-cgroup/containerd"

[debug]

address = "/data/labs/runtimes/docker-24/containerd-debug.sock"

format = ""

gid = 0

level = ""

uid = 0

[grpc]

address = "/data/labs/runtimes/docker-24/containerd-grpc.sock"

gid = 0

max_recv_message_size = 16777216

max_send_message_size = 16777216

tcp_address = "127.0.0.1:2375"

tcp_tls_ca = "/data/labs/runtimes/docker-24/ca.crt"

tcp_tls_cert = "/data/labs/runtimes/docker-24/tls.crt"

tcp_tls_key = "/data/labs/runtimes/docker-24/tls.key"

uid = 0

[metrics]

address = "127.0.0.1:1338"

grpc_histogram = false

[plugins]

[plugins."io.containerd.gc.v1.scheduler"]

deletion_threshold = 0

mutation_threshold = 100

pause_threshold = 0.02

schedule_delay = "0s"

startup_delay = "100ms"

[plugins."io.containerd.grpc.v1.cri"]

cdi_spec_dirs = ["/data/labs/runtimes/docker-24/cdi-cfgs", "/data/labs/runtimes/docker-24/cdi-run"]

device_ownership_from_security_context = false

disable_apparmor = false

disable_cgroup = false

disable_hugetlb_controller = true

disable_proc_mount = false

disable_tcp_service = true

drain_exec_sync_io_timeout = "0s"

enable_cdi = false

enable_selinux = false

enable_tls_streaming = false

enable_unprivileged_icmp = false

enable_unprivileged_ports = false

ignore_image_defined_volumes = false

image_pull_progress_timeout = "5m0s"

max_concurrent_downloads = 3

max_container_log_line_size = 16384

netns_mounts_under_state_dir = false

restrict_oom_score_adj = false

sandbox_image = "registry.k8s.io/pause:3.8"

selinux_category_range = 1024

stats_collect_period = 10

stream_idle_timeout = "4h0m0s"

stream_server_address = "127.0.0.1"

stream_server_port = "0"

systemd_cgroup = false

tolerate_missing_hugetlb_controller = true

unset_seccomp_profile = ""

disable_rdt = true

[plugins."io.containerd.grpc.v1.cri".cni]

bin_dir = "/data/labs/runtimes/docker-24/cni/bin"

conf_dir = "/data/labs/runtimes/docker-24/cni/net.d"

conf_template = ""

ip_pref = ""

max_conf_num = 1

setup_serially = false

[plugins."io.containerd.grpc.v1.cri".containerd]

default_runtime_name = "runc"

disable_snapshot_annotations = true

discard_unpacked_layers = false

ignore_blockio_not_enabled_errors = false

ignore_rdt_not_enabled_errors = false

no_pivot = false

snapshotter = "overlayfs"

address = "/data/labs/runtimes/docker-24/containerd.sock"

[plugins."io.containerd.grpc.v1.cri".containerd.default_runtime]

base_runtime_spec = ""

cni_conf_dir = "/data/labs/runtimes/docker-24/cni/net.d"

cni_max_conf_num = 0

container_annotations = []

pod_annotations = []

privileged_without_host_devices = false

privileged_without_host_devices_all_devices_allowed = false

runtime_engine = ""

runtime_path = "/data/labs/runtimes/docker-24/runc"

runtime_root = "/data/labs/runtimes/docker-24/runc-root"

runtime_type = ""

sandbox_mode = ""

snapshotter = ""

[plugins."io.containerd.grpc.v1.cri".containerd.default_runtime.options]

[plugins."io.containerd.grpc.v1.cri".containerd.runtimes]

[plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runc]

base_runtime_spec = ""

cni_conf_dir = "/data/labs/runtimes/docker-24/cni/net.d"

cni_max_conf_num = 0

container_annotations = []

pod_annotations = []

privileged_without_host_devices = false

privileged_without_host_devices_all_devices_allowed = false

runtime_engine = ""

runtime_path = "/data/labs/runtimes/docker-24/runc"

runtime_root = ""

runtime_type = "io.containerd.runc.v2"

sandbox_mode = "podsandbox"

snapshotter = ""

[plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runc.options]

BinaryName = ""

CriuImagePath = "/data/labs/runtimes/docker-24/criu-image"

CriuPath = ""

CriuWorkPath = "/data/labs/runtimes/docker-24/criu-work"

IoGid = 0

IoUid = 0

NoNewKeyring = false

NoPivotRoot = false

Root = ""

ShimCgroup = ""

SystemdCgroup = false

[plugins."io.containerd.grpc.v1.cri".containerd.untrusted_workload_runtime]

base_runtime_spec = ""

cni_conf_dir = "/data/labs/runtimes/docker-24/cni/net.d"

cni_max_conf_num = 0

container_annotations = []

pod_annotations = []

privileged_without_host_devices = false

privileged_without_host_devices_all_devices_allowed = false

runtime_engine = ""

runtime_path = "/data/labs/runtimes/docker-24/runc"

runtime_root = "/data/labs/runtimes/docker-24/runc-root"

runtime_type = ""

sandbox_mode = ""

snapshotter = ""

[plugins."io.containerd.grpc.v1.cri".containerd.untrusted_workload_runtime.options]

[plugins."io.containerd.grpc.v1.cri".image_decryption]

key_model = "node"

[plugins."io.containerd.grpc.v1.cri".registry]

config_path = "/data/labs/runtimes/docker-24/cri-registry.json"

[plugins."io.containerd.grpc.v1.cri".registry.auths]

[plugins."io.containerd.grpc.v1.cri".registry.configs]

[plugins."io.containerd.grpc.v1.cri".registry.headers]

[plugins."io.containerd.grpc.v1.cri".registry.mirrors]

[plugins."io.containerd.grpc.v1.cri".x509_key_pair_streaming]

tls_cert_file = "/data/labs/runtimes/docker-24/tls.crt"

tls_key_file = "/data/labs/runtimes/docker-24/tls.key"

[plugins."io.containerd.internal.v1.opt"]

path = "/data/labs/runtimes/docker-24/containerd-options"

[plugins."io.containerd.internal.v1.restart"]

interval = "10s"

[plugins."io.containerd.internal.v1.tracing"]

sampling_ratio = 1.0

service_name = "containerd"

[plugins."io.containerd.metadata.v1.bolt"]

content_sharing_policy = "shared"

[plugins."io.containerd.monitor.v1.cgroups"]

no_prometheus = false

[plugins."io.containerd.nri.v1.nri"]

disable = true

disable_connections = false

plugin_config_path = "/data/labs/runtimes/docker-24/nri/conf.d"

plugin_path = "/data/labs/runtimes/docker-24/nri/plugins"

plugin_registration_timeout = "5s"

plugin_request_timeout = "2s"

socket_path = "/data/labs/runtimes/docker-24/nri/nri.sock"

# runc的版本应与当前docker版本相匹配

[plugins."io.containerd.runtime.v1.linux"]

no_shim = false

runtime = "/data/labs/runtimes/docker-24/runc"

runtime_root = "/data/labs/runtimes/docker-24/runc-runtime-root"

shim = "/data/labs/runtimes/docker-24/containerd-shim-runc-v2"

shim_debug = false

[plugins."io.containerd.runtime.v2.task"]

platforms = ["linux/amd64"]

sched_core = false

[plugins."io.containerd.service.v1.diff-service"]

default = ["walking"]

[plugins."io.containerd.service.v1.tasks-service"]

blockio_config_file = "/data/labs/runtimes/docker-24/blockio.toml"

rdt_config_file = "/data/labs/runtimes/docker-24/rdt.toml"

#[plugins."io.containerd.snapshotter.v1.aufs"]

# root_path = "/data/labs/runtimes/docker-24/aufs"

[plugins."io.containerd.snapshotter.v1.blockfile"]

fs_type = ""

mount_options = []

root_path = ""

scratch_file = ""

#[plugins."io.containerd.snapshotter.v1.devmapper"]

# disabled = true

# async_remove = false

# base_image_size = ""

# discard_blocks = false

# fs_options = ""

# fs_type = ""

# pool_name = ""

# root_path = ""

[plugins."io.containerd.snapshotter.v1.native"]

root_path = "/data/labs/runtimes/docker-24/native-root"

[plugins."io.containerd.snapshotter.v1.overlayfs"]

mount_options = []

root_path = "/data/labs/runtimes/docker-24/overlayfs-root"

sync_remove = false

upperdir_label = false

[plugins."io.containerd.snapshotter.v1.zfs"]

root_path = "/data/labs/runtimes/docker-24/zfs-root"

[plugins."io.containerd.tracing.processor.v1.otlp"]

endpoint = ""

insecure = false

protocol = ""

[plugins."io.containerd.transfer.v1.local"]

config_path = "/data/labs/runtimes/docker-24/transfer.json"

max_concurrent_downloads = 3

max_concurrent_uploaded_layers = 3

[[plugins."io.containerd.transfer.v1.local".unpack_config]]

differ = ""

platform = "linux/amd64"

snapshotter = "overlayfs"

[proxy_plugins]

[stream_processors]

[stream_processors."io.containerd.ocicrypt.decoder.v1.tar"]

accepts = ["application/vnd.oci.image.layer.v1.tar+encrypted"]

args = ["--decryption-keys-path", "/data/labs/runtimes/docker-24/containerd/ocicrypt/keys"]

env = ["OCICRYPT_KEYPROVIDER_CONFIG=/data/labs/runtimes/docker-24/containerd/ocicrypt/ocicrypt_keyprovider.conf"]

path = "ctd-decoder"

returns = "application/vnd.oci.image.layer.v1.tar"

[stream_processors."io.containerd.ocicrypt.decoder.v1.tar.gzip"]

accepts = ["application/vnd.oci.image.layer.v1.tar+gzip+encrypted"]

args = ["--decryption-keys-path", "/data/labs/runtimes/docker-24/containerd/ocicrypt/keys"]

env = ["OCICRYPT_KEYPROVIDER_CONFIG=/data/labs/runtimes/docker-24/containerd/ocicrypt/ocicrypt_keyprovider.conf"]

path = "ctd-decoder"

returns = "application/vnd.oci.image.layer.v1.tar+gzip"

[timeouts]

"io.containerd.timeout.bolt.open" = "0s"

"io.containerd.timeout.metrics.shimstats" = "2s"

"io.containerd.timeout.shim.cleanup" = "5s"

"io.containerd.timeout.shim.load" = "5s"

"io.containerd.timeout.shim.shutdown" = "3s"

"io.containerd.timeout.task.state" = "2s"

[ttrpc]

address = "/data/labs/runtimes/docker-24/containerd-ttrpc.sock"

gid = 0

uid = 0

1.1.2 containerd的启动

【启动脚本】

>>> start-containerd.sh <<<

#!/bin/bash

nohup ./containerd --config $(pwd)/containerd-config.toml 1>>nohup-containerd.log 2>>nohup-containerd.log &

echo $! > containerd.pid

【停止脚本】

>>> stop-containerd.sh <<<

#!/bin/bash

pid=$(cat containerd.pid)

if [ ! -z "$pid" ]; then

kill -9 $pid 2>/dev/null

1.2 dockerd环境构建

1.2.1 为dockerd创建新的网桥（隔离的必要措施）

brctl show

brctl addbr docker1

ip link set dev docker1 up

ip addr add 172.24.0.0/24 dev docker1

ip addr show docker1

1.2.2 dockerd配置文件

【注意】在dockerd应指向containerd的GRPC协议socket地址，以进行通信和协作；其他路径应与原有dockerd环境相互隔离，避免相互影响。

>>> daemon.json <<<

{

"bridge": "docker1",

"containerd": "/data/labs/runtimes/docker-24/containerd-grpc.sock",

"data-root": "/data/labs/runtimes/docker-24/docker-data",

"debug": false,

"default-runtime": "runc-1.1.12",

"runtimes": {

"runc-1.1.12": {

"path": "/data/labs/runtimes/docker-24/runc",

"runtimeArgs": []

}

"exec-opts": ["native.cgroupdriver=cgroupfs", "proxy.path=/data/labs/runtimes/docker-24/docker-proxy"],

"exec-root": "/data/labs/runtimes/docker-24/docker-exec",

"host": [

"unix:///data/labs/runtimes/docker-24/docker.sock"

"init-path": "/data/labs/runtimes/docker-24/docker-init",

"pidfile": "/data/labs/runtimes/docker-24/docker.pid",

"proxies": {

"http-proxy": "http://192.168.0.100:7890",

"https-proxy": "http://192.168.0.100:7890",

"no-proxy": "localhost,127.0.0.1"

"log-driver": "json-file",

"log-level": "error",

"log-opts": {

"max-size": "10m",

"max-file": "3"

"max-concurrent-downloads": 10,

"storage-driver": "overlay2",

"userland-proxy": false

}

1.2.3 dockerd的启动

【启动脚本】

>>> start-dockerd.sh <<<

#!/bin/bash

nohup ./dockerd --config-file $(pwd)/daemon.json 1>>$(pwd)/dockerd.log 2>>$(pwd)/dockerd.log &

echo $! > dockerd.pid

【停止脚本】

>>> stop-dockerd.sh <<<

#!/bin/bash

pid=$(cat dockerd.pid)

if [ ! -z "$pid" ]; then

kill -9 $pid 2>/dev/null

1.3 ragflow搭建

一、下载ragflow源码：git clone https://github.com/infiniflow/ragflow.git

【注意】当前ragflow版本为v0.12.0

二、修改docker-compose配置清单，在ragflow/docker目录下

【注意】ragflow开箱自带了通义千问的云服务，但无法通过代理访问，因此无论在docker-compose.yml还是.env中设置代理均无效。由于GPU资源充足，后续将本地部署qwen2.5:72b模型，用于知识库检索增强。

三、修改数据库初始化脚本init.sql

>>> init.sql <<<

CREATE DATABASE IF NOT EXISTS rag_flow;

USE rag_flow;

-- 添加以下内容：设置 root 用户允许从任何 IP 地址连接，因为ragflow-server通过容器私有地址访问ragflow-mysql，这个容器的IP是不确定的

ALTER USER 'root'@'localhost' IDENTIFIED BY 'infini_rag_flow';

CREATE USER 'root'@'%' IDENTIFIED BY 'infini_rag_flow';

GRANT ALL PRIVILEGES ON *.* TO 'root'@'%' WITH GRANT OPTION;

FLUSH PRIVILEGES;

【注意】若mysql初始化失败，再次尝试应首先删除mysql数据目录，位置在：${docker-data}/volumes/docker_mysql_data/_data下。

四、（可选步骤）为了避免与服务器其他容器产生冲突，我们将所有容器到宿主机的映射端口关闭，仅保留必要的ragflow-server的前端开放端口

五、调整mysql的卷挂载路径，修改容器内路径为/init.sql

mysql:

volumes:

#- ./init.sql:/data/application/init.sql:ro

- ./init.sql:/init.sql:ro

【注意】本人在测试时发现，可能因mysql容器默认账户为非root，导致/data/application/init.sql路径挂载失败，调整为/init.sql运行正常

六、使用新的docker运行环境启动ragflow容器

DOCKER_HOST=unix:///data/labs/runtimes/docker-24/docker.sock MACOS= /data/labs/runtimes/docker-24/docker-compose -f docker-compose.yml up -d

1.4 本地LLM部署

这里使用ollama部署qwen2.5:72b模型，大约需要57G显存。

一、为使得ollama可以通过代理更新模型包，可以在/etc/systemd/system/ollama.service中增加以下配置：

[Service]

Environment="OLLAMA_HOST=<GPU主机IP>:8088"

Environment="OLLAMA_MODELS=/data/labs/ollama"

Environment="HTTP_PROXY=http://192.168.0.100:7890"

Environment="HTTPS_PROXY=http://192.168.0.100:7890"

二、下载和运行

OLLAMA_HOST=192.168.0.101:8088 ollama pull qwen2.5:72b

OLLAMA_HOST=192.168.0.101:8088 ollama run qwen2.5:72b

三、ollama后台运行LLM，目前没有找到解决方案，而采用了tmux终端复用方案，具体命令为：

【创建会话】 tmux new -s ollama-session

【退出会话】 Ctrl+b 再按d键

【查询会话】 tmux list-sessions

【重进会话】 tmux attach -t ollama-session

2 ragflow体验

注册与登录完成后即可开始体验。

2.1 配置知识库

默认项也许是最优项，这里的选项基本可以不用动了。需要注意的是，在 RAG系统中，嵌入模型（Embedding Model）是用于将文本转化为向量表示的模型。这些向量表示（也称为嵌入）能够捕捉文本的语义信息，使得相似的文本在向量空间中距离较近，从而便于进行相似性计算和检索。这里嵌入模型选择bge-large-zh-v1.5的理由是，其在生成高质量的文本嵌入方面具有优势。

文本嵌入（Text Embedding）是将文本数据转化为固定维度向量表示的过程，这些向量可以捕捉文本的语义信息，使相似的文本在向量空间中距离较近，从而便于进行相似度计算、分类和聚类任务。举例说明：1）我们有个新闻文章数据库，每篇文章都有一个标题和内容；2）生成嵌入：即使用预训练的嵌入模型将每篇文章的内容转化为向量表示；3）查询嵌入：用户输入一个查询，我们将查询文本转化为向量表示；4）计算相似度：计算查询向量与每篇文章向量的相似度，找到最相似的文章。

2.2 上传一个Excel知识库文档并解析

这里以“系统漏洞列表”Excel做为示例演示，约66000行数据，用了大约3个小时解析完成。

2.3 在模型提供商中添加ollama语言模型

2.4 创建聊天助理

2.5 使用不同嵌入模型的效果比较

知识库原始数据部分条目如下：（敏感信息已隐藏）

安全管理	系统安全管理	季度	系统满足各项安全管理要求，不符合标准，扣2分/次；
	人员安全管理	季度	人员管理使用满足各项安全管理要求，包括人员出入、账号管理等，不符合标准，扣2分/次；
	安全事故	季度	由于**团队运维人员人为原因造成1次安全事故，导致系统故障、被或**通报，每次扣5分；