English | 简体中文 | Help

RAG Has Never Been So Easy

minRAG is a RAG system built from scratch, pursuing extreme simplicity and power, with no more than 10,000 lines of code. It supports AI platforms such as OpenAI, Gitee AI, Baidu Qianfan, Tencent Cloud LKE, Alibaba Cloud BaiLian, and ByteDance Volcano Engine. No installation is required; just double-click to start.
It uses FTS5 for BM25 full-text search and Vec for vector retrieval. It implements components such as MarkdownConverter, DocumentSplitter, OpenAIDocumentEmbedder, SQLiteVecDocumentStore, OpenAITextEmbedder, VecEmbeddingRetriever, FtsKeywordRetriever, DocumentChunkReranker, PromptBuilder, OpenAIChatMemory, OpenAIChatGenerator, ChatMessageLogStore, Pipeline, etc., supporting pipeline setup and extension.

Supported AI PlatformsSince

reranker does not have a unified standard, the base_url in the component parameters must be filled in with the complete path.

OpenAI

minRAG implements the standard specifications of OpenAI, and all platforms compatible with OpenAI can be used.

Gitee AI (Default)

The default AI platform is Gitee AI. Gitee AI offers 100 free calls per day.

• On the registration or settings page, fill in base_url for the AI platform as https://ai.gitee.com/v1.
• On the registration or settings page, fill in api_key with a free or paid token.
• OpenAITextEmbedder defaults to using the bge-m3 model.
• The component parameter for GiteeDocumentChunkReranker is {"base_url":"https://ai.gitee.com/api/serverless/bge-reranker-v2-m3/rerank","model":"bge-reranker-v2-m3"}.
• OpenAIChatGenerator is recommended to use the DeepSeek-V3.2-Exp model.

Tencent Cloud LKE Knowledge Engine

• On the registration or settings page, fill in base_url for the AI platform as SecretId, or configure it in the component parameters as {"SecretId":"xxx"}.
• On the registration or settings page, fill in api_key as SecretKey, or configure it in the component parameters as {"SecretKey":"xxx"}.
• LKETextEmbedder and LKEDocumentEmbedder default to using the lke-text-embedding-v1 model.
• LKEDocumentChunkReranker defaults to using the lke-reranker-base model.
• OpenAIChatGenerator connects using the OpenAI SDK method. The component parameter configuration is {"base_url":"https://api.lkeap.cloud.tencent.com/v1","api_key":"xxx","model":"deepseek-v3.2-exp"}.
• Remember to modify the components in the pipeline.

Baidu QianFan

• On the registration or settings page, fill in base_url for the AI platform as https://qianfan.baidubce.com/v2.
• On the registration or settings page, fill in api_key with a permanently valid API Key.
• OpenAITextEmbedder and OpenAIDocumentEmbedder default to using the bge-large-zh model, with 1024 dimensions.
• The component parameter configuration for DocumentChunkReranker is {"base_url":"https://qianfan.baidubce.com/v2/rerankers","model":"bce-reranker-base","top_n":5,"score":0.1}.
• OpenAIChatGenerator is recommended to use the deepseek-v3.2 model.
• Remember to modify the components in the pipeline.

Alibaba Cloud BaiLian

• On the registration or settings page, fill in base_url for the AI platform as https://dashscope.aliyuncs.com/compatible-mode/v1.
• On the registration or settings page, fill in api_key with the applied API KEY.
• OpenAITextEmbedder and OpenAIDocumentEmbedder default to using the text-embedding-v3 model, with 1024 dimensions.
• The component parameter configuration for BaiLianDocumentChunkReranker is {"base_url":"https://dashscope.aliyuncs.com/api/v1/services/rerank/text-rerank/text-rerank","model":"qwen3-rerank","top_n":5,"score":0.1}.
• OpenAIChatGenerator is recommended to use the deepseek-v3.2-exp model.
• Remember to modify the components in the pipeline.

ByteDance Volcano Engine

• On the registration or settings page, fill in base_url for the AI platform as https://ark.cn-beijing.volces.com/api/v3.
• On the registration or settings page, fill in api_key with the applied API KEY.
• OpenAITextEmbedder and OpenAIDocumentEmbedder are recommended to use the doubao-embedding model, compatible with 1024 dimensions.
• Volcano Engine does not currently have a Reranker model, so it is recommended to use Reranker models from other platforms or remove it.
• OpenAIChatGenerator is recommended to use the deepseek-v3-1-terminus model.
• Remember to modify the components in the pipeline.

UI Preview

Markitdown Integration (Default)

Use https://gitee.com/minrag/markitdown to parse documents, The compiled dist/markitdown from python build.py should be placed in the minragdatadir directory. Example configuration for the MarkdownConverter component:

{
    // Image analysis model
    "model":"Qwen3-VL-30B-A3B-Instruct", 
    // Understand the prompts for images in the document
    "prompt":"Accurately extract the content of the image and directly describe the image without any irrelevant guidance or similar information.", 
    // Markdown command path
    "markitdown":"minragdatadir/markitdown/markitdown",
    // Generated Markdown File Directory
    "markdownDir":"minragdatadir/upload/markitdown/markdown",
    // Directory for storing images
    "imageFileDir":"minragdatadir/upload/markitdown/image",
    // URL prefix directory
    "imageURLDir":"/upload/markitdown/image"
}

Tika Integration

By default, minRAG only supports text formats like Markdown and plain text. To parse document content, you can use the TikaConverter component to call the tika service. Example configuration for TikaConverter:

{
    "tikaURL": "http://localhost:9998/tika",
    "defaultHeaders": {
        "Content-Type": "application/octet-stream"
    }
}

Start the tika service with this command:

## Apache Tika 3.x requires JDK 11 or later
java -jar tika-server-standard-3.1.0.jar --host=0.0.0.0 --port=9998

## To disable logging
#nohup java -jar tika-server-standard-3.1.0.jar --host=0.0.0.0 --port=9998 >/dev/null 2>&1 &

Alternatively, download tika-windows and use start.bat to launch Tika.

Note: Modify the parameters in your indexPipeline workflow by replacing the original MarkdownConverter with TikaConverter:

{
    "start": "TikaConverter",
    "process": {
        "TikaConverter": "DocumentSplitter",
        "DocumentSplitter": "OpenAIDocumentEmbedder",
        "OpenAIDocumentEmbedder": "SQLiteVecDocumentStore"
    }
}

Development Environment

minRAG uses https://github.com/wangfenjin/simple as the FTS5 full-text search extension. The compiled libsimple file is placed in the minragdatadir/extensions directory. If minRAG fails to start and reports an error connecting to the database, please check if the libsimple file is correct. If you need to recompile libsimple, please refer to https://github.com/wangfenjin/simple.

The default port is 738, and the backend management address is http://127.0.0.1:738/admin/login.
First, unzip minragdatadir/dict.zip.
Run go run --tags "fts5" ..
Package: go build --tags "fts5" -ldflags "-w -s".

The development environment requires CGO compilation configuration. Set set CGO_ENABLED=1, download mingw64 and cmake, and configure the bin to the environment variables. Note to rename mingw64/bin/mingw32-make.exe to make.exe.
Modify vscode's launch.json to add ,"buildFlags": "--tags=fts5" for debugging fts5.
Test needs to be done manually: go test -v -count=1 -timeout 30s --tags "fts5" -run ^TestVecQuery$ gitee.com/minrag/minrag.
Package: go build --tags "fts5" -ldflags "-w -s".
When recompiling simple, it is recommended to use the precompiled version from https://github.com/wangfenjin/simple.
Note to modify the Windows compilation script, remove the libgcc_s_seh-1.dll and libstdc++-6.dll dependencies for mingw64 compilation, and turn off BUILD_TEST_EXAMPLE as there are conflicts.

rmdir /q /s build
mkdir build && cd build
cmake .. -G "Unix Makefiles" -DBUILD_TEST_EXAMPLE=OFF -DCMAKE_INSTALL_PREFIX=release -DCMAKE_CXX_FLAGS="-static-libgcc -static-libstdc++" -DCMAKE_EXE_LINKER_FLAGS="-Wl,-Bstatic -lstdc++ -lpthread -Wl,-Bdynamic"
make && make install

Backend Management Supports English

The minRAG backend management currently supports both Chinese and English, with the capability to extend to other languages. Language files are located in minragdatadir/locales. By default, the system uses Chinese (zh-CN) upon initial installation. If English is preferred, you can modify the "locale":"zh-CN" to "locale":"en-US" in the minragdatadir/install_config.json file before installation. Alternatively, after successful installation, you can change the Language setting to English in the Settings and restart the system to apply the changes.

Table Structure

ID defaults to timestamp (23 digits) + random number (9 digits), globally unique.
Table creation statement minragdatadir/minrag.sql

Configuration (Table Name: config)

Reads minragdatadir/install_config.json during installation.

columnName	Type	Description	Remarks
id	string	Primary Key	minrag_config
basePath	string	Base Path	Default /
jwtSecret	string	JWT Secret Key	Randomly generated
jwttokenKey	string	JWT Key	Default jwttoken
serverPort	string	IP:Port	Default :738
timeout	int	JWT Timeout in seconds	Default 7200
maxRequestBodySize	int	Maximum Request Body Size	Default 20M
locale	string	Language Pack	Default zh-CN,en-US
proxy	string	HTTP Proxy Address
createTime	string	Creation Time	2006-01-02 15:04:05
updateTime	string	Update Time	2006-01-02 15:04:05
createUser	string	Creator	Initialized as system
sortNo	int	Sort Order	Descending
status	int	Status	Disabled(0), Enabled(1)

User (Table Name: user)

columnName	Type	Description	Remarks
id	string	Primary Key	minrag_admin
account	string	Login Name	Default admin
passWord	string	Password	-
userName	string	Description	-
createTime	string	Creation Time	2006-01-02 15:04:05
updateTime	string	Update Time	2006-01-02 15:04:05
createUser	string	Creator	Initialized as system
sortNo	int	Sort Order	Descending
status	int	Status	Disabled(0), Enabled(1)

Site Information (Table Name: site)

columnName	Type	Description	Remarks
id	string	Primary Key	minrag_site
title	string	Site Name	-
keyword	string	Keywords	-
description	string	Site Description	-
theme	string	Default Theme	Default is default
themePC	string	PC Theme	Fetched from cookie first, if not, from Header, then written to cookie, default is default
themeWAP	string	Mobile Theme	Fetched from cookie first, if not, from Header, then written to cookie, default is default
themeWX	string	WeChat Theme	Fetched from cookie first, if not, from Header, then written to cookie, default is default
logo	string	Logo	-
favicon	string	Favicon	-
createTime	string	Creation Time	2006-01-02 15:04:05
updateTime	string	Update Time	2006-01-02 15:04:05
createUser	string	Creator	Initialized as system
sortNo	int	Sort Order	Descending
status	int	Status	Disabled(0), Enabled(1)

Knowledge Base (Table Name: knowledgeBase)

columnName	Type	Description	Remarks
id	string	Primary Key	URL path, separated by /, e.g., /web/
name	string	Knowledge Base Name	-
pid	string	Parent Knowledge Base ID	Parent Knowledge Base ID
knowledgeBaseType	int	Knowledge Base Type	-
createTime	string	Creation Time	2006-01-02 15:04:05
updateTime	string	Update Time	2006-01-02 15:04:05
createUser	string	Creator	Initialized as system
sortNo	int	Sort Order	Descending
status	int	Status	Disabled(0), Enabled(1)

Document (Table Name: document)

columnName	Type	Description	Tokenized	Remarks
id	string	Primary Key	No	URL path, separated by /, e.g., /web/nginx-use-hsts
name	string	Document Name	No	-
knowledgeBaseID	string	Knowledge Base ID	No	-
knowledgeBaseName	string	Knowledge Base Name	No	-
toc	string	Table of Contents	No	-
summary	string	Summary	No	-
markdown	string	Markdown Content	No	-
filePath	string	File Path	No	-
fileSize	int	File Size	No	-
fileExt	string	File Extension	No	-
createTime	string	Creation Time	-	2006-01-02 15:04:05
updateTime	string	Update Time	-	2006-01-02 15:04:05
createUser	string	Creator	-	Initialized as system
sortNo	int	Sort Order	-	Descending
status	int	Status	-	Disabled(0), Enabled(1), Processing(2), Failed(3)

Document Chunk (Table Name: document_chunk)

columnName	Type	Description	Tokenized	Remarks
id	string	Primary Key	No	-
documentID	string	Document ID	No	-
knowledgeBaseID	string	Knowledge Base ID	No	-
knowledgeBaseName	string	Knowledge Base Name	No	-
markdown	string	Markdown Content	Yes	Using jieba tokenizer
createTime	string	Creation Time	-	2006-01-02 15:04:05
updateTime	string	Update Time	-	2006-01-02 15:04:05
createUser	string	Creator	-	Initialized as system
sortNo	int	Sort Order	-	Descending
status	int	Status	-	Disabled(0), Enabled(1), Processing(2), Failed(3)

Component (Table Name: component)

columnName	Type	Description	Tokenized	Remarks
id	string	Primary Key	No	-
componentType	string	Component Type	No	-
parameter	string	Component Parameters	No	-
createTime	string	Creation Time	-	2006-01-02 15:04:05
updateTime	string	Update Time	-	2006-01-02 15:04:05
createUser	string	Creator	-	Initialized as system
sortNo	int	Sort Order	-	Descending
status	int	Status	-	Disabled(0), Enabled(1)

Agent (Table Name: agent)

columnName	Type	Description	Tokenized	Remarks
id	string	Primary Key	No	-
name	string	Agent Name	No	-
knowledgeBaseID	string	Knowledge Base ID	No	-
pipelineID	string	Pipeline ID	No	-
defaultReply	string	Default Reply	No	-
agentType	int	Agent Type	No	-
agentPrompt	string	Agent Prompt	No	-
avatar	string	Agent Avatar	No	-
welcome	string	Welcome Message	No	-
tools	string	Functions to Call	No	-
memoryLength	int	Context Memory Length	No	-
createTime	string	Creation Time	-	2006-01-02 15:04:05
updateTime	string	Update Time	-	2006-01-02 15:04:05
createUser	string	Creator	-	Initialized as system
sortNo	int	Sort Order	-	Descending
status	int	Status	-	Disabled(0), Enabled(1)

Chat Room (Table Name: chat_room)

Column Name	Type	Description	Tokenized	Remarks
id	string	Primary Key	No	-
name	string	Chat Room Name	No	-
agentID	string	Agent ID	No	-
pipelineID	string	Pipeline ID	No	-
knowledgeBaseID	string	Knowledge Base ID	No	-
userID	string	User ID	No	-
createTime	string	Creation Time	No	2006-01-02 15:04:05

Message Log (Table Name: message_log)

Column Name	Type	Description	Tokenized	Remarks
id	string	Primary Key	No	-
agentID	string	Agent ID	No	-
roomID	string	Chat Room ID	No	-
pipelineID	string	Pipeline ID	No	-
knowledgeBaseID	string	Knowledge Base ID	No	-
userMessage	string	User's Message	No	-
aiMessage	string	AI's Response	No	-
userID	string	User ID	No	-
createTime	string	Creation Time	No	2006-01-02 15:04:05

Copyright and Software Copyright Description

• The software copyright registration number of this minRAG is 2025SR0616004
• The software copyright of this minRAG is owned by us, and secondary software copyright applications are prohibited. Infringement will be prosecuted
• The copyright of programs developed by developers using minRAG belongs to the developers
• Please retain the copyright without any other restrictions. That is to say, you must include the original license agreement statement in your distribution, whether you distribute it in binary or source code form
• The open-source version is released under the AGPL-3.0 open-source license and is provided for free use, but it is not allowed to release and sell modified and derivative code as closed-source commercial software!