Fork: 112 Star: 1265 (更新于 2024-05-29 10:21:02)

license: MIT

Language: HTML .

The framework for fast development and deployment of RAG systems.

最后发布版本: v0.2.00 ( 2024-05-23 17:52:49)

官方网址 GitHub网址

Docs Discord Github Stars Commits-per-week License: MIT

Sciphi Framework

Build, deploy, observe, and optimize your RAG system.


R2R, short for RAG to Riches, provides the fastest and most efficient way to deliver high-quality Retrieval-Augmented Generation (RAG) to end users. The framework is built around customizable pipelines and a feature-rich FastAPI implementation.


R2R was conceived to bridge the gap between local LLM experimentation and scalable production solutions. It is built with observability and customization in mind, ensuring that users can seamlessly transition from development to deployment.

Key Features

  • 🔧 Build: Use the framework to build arbitrary asynchronous pipelines.
  • 🚀 Deploy: Instantly launch production-ready asynchronous RAG pipelines with streaming capabilities.
  • 🧩 Customize: Tailor your multimodal pipeline with intuitive configuration files.
  • 🔌 Extend: Enhance your pipeline with custom code integrations.
  • 🤖 OSS: Benefit from a framework developed by the open-source community, designed to simplify RAG deployment.

Table of Contents

  1. Quick Install
  2. Links
  3. R2R Demo
  4. R2R Server-Client Demo
  5. Core Abstractions
  6. Summary

Quick Install:

# use the `'r2r[all]'` to download all required deps
pip install r2r

# setup env 
export OPENAI_API_KEY=sk-...


Join the Discord server

R2R Docs Quickstart

SciPhi Cloud Docs

Local RAG Tutorial

R2R Demo

The R2R demo offers a step-by-step guide on running the default R2R Retrieval-Augmented Generation (RAG) pipeline. The demo ingests a list of provided provided documents and demonstrates search, RAG, and advanced functionality. The script at r2r/examples/, which powers the demo, can be configured and extended with sufficient developer familiarity.

Ingest Demo Files

To comprehensively demonstrate the RAG functionalities of the R2R framework, we must start by ingesting a realistic set of documents. Running the command below will parse, chunk, embed, and store a preset list of files. The included file types cover HTML, PDF, PNG, and TXT examples:

python -m r2r.examples.demo ingest_as_files

Demo Output:

r2r.pipes.parsing_pipe - INFO - Parsed document with metadata={'title': 'pg_essay_1.html', 'user_id': '063edaf8-3e63-4cb9-a4d6-a855f36376c3'} and id=4a4fb848-fc03-5487-a7e5-33c9fdfb73cc in t=0.00 seconds. - 2024-05-21 08:39:59,003
r2r.pipes.parsing_pipe - INFO - Parsed document with metadata={'title': 'lyft_2021.pdf', 'user_id': '063edaf8-3e63-4cb9-a4d6-a855f36376c3'} and id=c5abc0b7-b9e5-54d9-b3d3-fdb14af4d065 in t=3.47 seconds. - 2024-05-21 08:40:02,477
r2r.pipes.parsing_pipe - INFO - Parsed document with metadata={'title': 'screen_shot.png', 'user_id': '063edaf8-3e63-4cb9-a4d6-a855f36376c3', 'image_type': 'png'} and id=74f1506a-9a37-59d7-b288-5ef3683dca8f in t=18.37 seconds. - 2024-05-21 08:40:32,310
Time taken to ingest files: 28.49 seconds

Confirm User Data

To verify the successful ingestion of the demo documents, you can fetch the metadata for the uploaded documents associated with the default demo user ID:

python -m r2r.examples.demo get_user_document_data --user_id="063edaf8-3e63-4cb9-a4d6-a855f36376c3"

Demo Output:

Time taken to get user document data: 0.21 seconds
         'document_id': '327f6110-edd1-5fe3-b6b3-49b55f1cbc28',
         'title': 'pg_essay_3.html'
         'document_id': '946859f0-da5c-5db7-9b5c-c586be76d709', 
         'title': 'pg_essay_5.html'
         'document_id': '64c1c913-be06-548f-acbc-3618b00d3616', 
         'title': 'lyft_2021.pdf'

Search Demo Documents

Documents are stored by default in a local vector database. The vector database provider and settings can be specified via an input config.json. To perform a search query on the ingested user documents, use the following command:

python -m r2r.examples.demo search --query="Who was Aristotle?"

Demo Output:

Time taken to search: 0.39 seconds
   'id': UUID('93c44e73-8e95-50c2-84af-6a42f070b552'), 
   'score': 0.7739712385010018, 
      'document_id': '15255e98-e245-5b58-a57f-6c51babf72dd', 
      'extraction_id': '5c61f9b9-b468-5fd7-8eb1-5d797a15c484', 
      'text': 'Aristotle[A] (Greek: Ἀριστοτέλης Aristotélēs, pronounced [aristotélɛːs]; 384–322 BC) was an Ancient Greek philosopher and polymath. His writings cover a broad range of subjects spanning the natural sciences, philosophy, linguistics, economics, politics,

 psychology, and the arts. As the founder of the Peripatetic school of philosophy in the Lyceum in Athens, he began the wider Aristotelian tradition that followed, which set the groundwork for the development of modern science.', 
      'title': 'aristotle.txt',
      'user_id': '063edaf8-3e63-4cb9-a4d6-a855f36376c3', 
      'query': 'Who was Aristotle?'

RAG Demo

Completion Response

To generate a response for a query using RAG, execute the following command:

python -m r2r.examples.demo rag --query="What was Uber's profit in 2020?"

Demo Output:

Time taken to run RAG: 2.29 seconds
                  content="Uber's profit in 2020 was a net loss of $6,768 million [10].", 

Streaming Response

For streaming results from a RAG query, use the following command:

python -m r2r.examples.demo rag --query="What was Lyft's profit in 2020?" --streaming=True

Demo Output:

r2r.main.r2r_config - INFO - Loading configuration from <YOUR_WORKDIR>/config.json - 2024-05-20 22:27:31,890
<search>["{\"id\":\"808c47c5-ebef-504a-a230-aa9ddcfbd87 .... </search>
<completion>Lyft reported a net loss of $1,752,857,000 in 2020 according to [2]. Therefore, Lyft did not make a profit in 2020.</completion>                                                      
Time taken to stream RAG response: 2.79 seconds

Document Management Demo

Document Deletion

To delete a document by its ID, or any other metadata field, use the delete command. For example, to delete all chunks corresponding to the uploaded file aristotle.txt, we can call delete on the associated document ID with the value 15255e98-e245-5b58-a57f-6c51babf72dd:

python -m r2r.examples.demo delete --key=document_id --value=15255e98-e245-5b58-a57f-6c51babf72dd

User Deletion

To delete all documents associated with a given user, run the delete command on the user_id:

run the following command with care, as it will erase all ingested user data
python -m r2r.examples.demo delete --key=user_id --value=063edaf8-3e63-4cb9-a4d6-a855f36376c3

R2R Server-Client Demo

This section extends the previous demo by showing how to set up and use the R2R framework with a server-client architecture. The R2R server can be stood up to handle requests, while the client can communicate with the server to perform various operations.

Launch the Server

Use the following command to start the server:

python -m r2r.examples.demo serve

This command starts the R2R server on the default host and port 8000.

Example Commands

  1. Ingest Documents as Files:

    python -m r2r.examples.demo ingest_as_files --base_url=http://localhost:8000

    This command will send the ingestion request to the server running at http://localhost:8000.

  2. Perform a Search:

    python -m r2r.examples.demo search --query="Who was Aristotle?" --base_url=http://localhost:8000

    This command sends the search query to the server and retrieves the results.

  3. Run a RAG Completion:

    python -m r2r.examples.demo rag --query="What was Uber's profit in 2020?" --base_url=http://localhost:8000

    This command sends the RAG query to the server and retrieves the generated response.

  4. Run a RAG Stream:

    python -m r2r.examples.demo rag --query="What was Lyft's profit in 2020?" --streaming=True --base_url=http://localhost:8000

    This command streams the RAG query results from the server.

Server-Client Summary

By using the server-client model, you can extend the basic R2R demo to support more scalable and modular deployments. The server handles requests and performs heavy computations, while clients can communicate with the server to perform ingestion, search, RAG, and other operations, as shown in the examples above. For detailed setup and basic functionality, refer back to the R2R Demo.

Core Abstractions

The framework revolves around three core abstractions: Providers, Pipes, and Pipelines.


Providers supply the necessary resources and capabilities to the pipes and pipelines. Key provider types include:

  • Vector Database Provider: Manages the storage and retrieval of vector embeddings. Examples include Qdrant, PGVector, and LocalVectorDB.
  • Embedding Provider: Converts text into vector embeddings. Supported providers include OpenAI, SentenceTransformers, and DummyEmbeddingProvider.
  • LLM Provider: Interfaces with large language models for text generation. Supported providers include OpenAI, LiteLLM, and LlamaCPP.
  • Prompt Provider: Manages prompts for various tasks.
  • Eval Provider: Evaluates the quality of generated responses.


Pipes represent individual steps in the data processing workflow. Each pipe performs a specific task, such as parsing, embedding, searching, or generating text. Pipes are designed to be composable and reusable within different pipelines.

  • Parsing Pipe: Extracts and structures data from various formats.
  • Embedding Pipe: Generates embeddings from text and stores them in a vector database.
  • Vector Storage Pipe: Handles the storage of embeddings in a vector database.
  • Search Pipe: Performs vector-based searches.
  • RAG Pipe: Integrates search results with language generation to produce responses.
  • Streaming RAG Pipe: Extends RAG functionality to support streaming responses.
  • Eval Pipe: Evaluates the quality of generated responses.


Pipelines are composed of multiple pipes arranged in a sequence. They manage the flow of data through the pipes, ensuring that each step is executed in the correct order. R2R supports several types of pipelines:

  • Ingestion Pipeline: Prepares and ingests documents, converting them into embeddings.
  • Embedding Pipeline: Manages the transformation of text into vector embeddings.
  • RAG Pipeline: Combines search and language generation to produce detailed responses.
  • Eval Pipeline: Evaluates the quality of generated responses using LLM-powered evaluations.


R2R (RAG to Riches) is a comprehensive framework designed to streamline the development, deployment, and optimization of Retrieval-Augmented Generation (RAG) systems. With its robust core abstractions—Providers, Pipes, and Pipelines—R2R offers a modular and flexible approach to building high-quality RAG pipelines.

最近版本更新:(数据更新于 2024-05-29 10:20:47)

2024-05-23 17:52:49 v0.2.00

2024-04-17 08:53:13 v0.1.35

2024-04-10 14:23:41 v0.1.34

2024-04-10 14:00:00 v0.1.33

2024-04-10 13:26:21 v0.1.32

2024-04-06 05:53:24 v0.1.31

2024-04-05 01:32:36 v0.1.30

2024-04-03 14:32:02 v0.1.29

2024-03-29 13:10:33 v0.1.28

2024-03-26 00:27:16 v0.1.27


artificial-intelligence, chatbot, data-pipelines, deep-learning, langchain, large-language-models, llama-index, llm, machine-learning, ocr, pdf, question-answering, retrieval, retrieval-augmented-generation, retrieval-systems, search

SciPhi-AI/R2R同语言 HTML最近更新仓库

2024-07-18 00:44:48 windmill-labs/windmill

2024-07-14 19:52:20 squidfunk/mkdocs-material

2024-07-09 21:13:24 qiye45/wechatDownload

2024-07-09 19:33:57 Unstructured-IO/unstructured

2024-07-08 00:31:17 nova-video-player/aos-AVP

2024-07-05 16:04:07 Tencent/tdesign-miniprogram