Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/cohereai-labs/cers-subnet

Bittensor subnet focuses on optimizing and scaling Retrieval Augmented Generation (RAG) applications for enterprise use cases, leveraging Bittensor's decentralized compute to enhance Cohere's LLM performance and accuracy
https://github.com/cohereai-labs/cers-subnet

bittensor cohere

Last synced: 10 months ago
JSON representation

Bittensor subnet focuses on optimizing and scaling Retrieval Augmented Generation (RAG) applications for enterprise use cases, leveraging Bittensor's decentralized compute to enhance Cohere's LLM performance and accuracy

Awesome Lists containing this project

README 

          



# **Cohere Enterprise RAG Subnet xx** 

[![Discord Chat](https://img.shields.io/discord/308323056592486420.svg)](https://discord.gg/bittensor) 

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT) 



---



## Powering Secure, High-Performance Enterprise RAG 



[Discord](https://discord.gg/bittensor) • [@cohere](https://x.com/cohere) • [Website](https://www.cohere.com)




---

- [Introduction](#introduction)

- [What This Subnet Solves for an Enterprise](#what-this-subnet-solves-for-an-enterprise)

- [How it works on Bittensor](#how-it-works-on-bittensor)

- [🔒 Security and Data Privacy](#-security-and-data-privacy)

  - [Architecture Overview](#architecture-overview)

- [Getting Started](#getting-started)

  - [Prerequisites](#prerequisites)

  - [1. Clone \& Install](#1-clone--install)

  - [2. Configure Your Wallet](#2-configure-your-wallet)

  - [3. Run](#3-run)

    - [For Miners](#for-miners)

    - [For Validators](#for-validators)

- [License](#license)



---

## Introduction



Unlock the full potential of your enterprise data with the **Cohere Enterprise RAG Subnet**. In today's AI-driven landscape, leveraging proprietary information is the key to a competitive edge. However, building and scaling a secure, high-performance Retrieval-Augmented Generation (RAG) system presents a significant challenge for many organizations.



This Bittensor subnet provides a decentralized, incentivized solution. It allows enterprises to supercharge their AI applications with their own private data, tapping into a competitive network of retrieval models without ever compromising on security or performance.



## What This Subnet Solves for an Enterprise



By participating in the CERS subnet, businesses can overcome common RAG hurdles and gain a distinct advantage:



-   **🚀 Superior RAG Performance**: Move beyond generic retrieval models. Leverage a global network of miners who are incentivized to develop and operate highly specialized and optimized retrieval algorithms tailored to your specific domain and data types.



-   **🔒 Uncompromised Data Security**: Your proprietary documents and sensitive information **never** leave your secure environment. The subnet is designed with a privacy-first architecture that operates exclusively on non-reversible data embeddings. (See [Security and Data Privacy](#-security-and-data-privacy) for details).



-   **💰 Cost-Effective Scalability**: Offload the heavy computational and maintenance burden of indexing and retrieval to a decentralized network. This allows you to pay for top-tier performance without the massive overhead of building and maintaining complex infrastructure in-house.



-   **🧩 Unmatched Customization**: The subnet fosters a competitive ecosystem where miners can specialize in different industries (e.g., finance, healthcare, legal) and data formats. This gives you access to a marketplace of tailored solutions that best fit your unique business needs.



-   **🤝 Powered by Cohere & Bittensor**: This subnet brings together Cohere's leadership in enterprise-grade AI and Large Language Models with Bittensor's robust, decentralized incentive network, creating a powerful and reliable foundation for your RAG applications.



---



## How it works on Bittensor



This subnet leverages Bittensor's core incentive mechanism to create a competitive marketplace for high-performance information retrieval. Here's how the components work together:



-   **Validators as Enterprise Proxies**: Validators act on behalf of an enterprise. They send query embeddings (derived from user questions) to the network of miners. Their primary role is to evaluate the quality and relevance of the miners' responses.



-   **Miners as Specialized Retrieval Engines**: Miners are the workhorses of the subnet. Their job is to:

    1.  Securely receive and index document embeddings provided by the enterprise via a private API.

    2.  Run sophisticated, high-speed search algorithms to find the most relevant document embeddings for a given query embedding.

    3.  Return a ranked list of document IDs to the validator.



-   **Incentivizing Performance with TAO**: The magic happens in the evaluation step.

    1.  Validators use a benchmark dataset (a set of queries with known relevant document IDs) to score the responses from each miner.

    2.  Performance is measured based on retrieval accuracy (e.g., precision, recall, MRR).

    3.  Based on these scores, validators set weights on the Bittensor network. Miners who consistently provide the most accurate and fastest results receive a larger share of the network's TAO rewards.



This continuous cycle of competition and reward drives miners to constantly innovate and optimize their retrieval models, ensuring that enterprises always have access to a state-of-the-art, decentralized retrieval system for their RAG applications.



---



## 🔒 Security and Data Privacy



In an enterprise context, data privacy is paramount. This subnet is designed with a **privacy-first architecture** where sensitive enterprise data is never exposed to the public network or the miners.



### Architecture Overview



The following diagram illustrates the data flow and the separation of concerns between the enterprise environment and the Bittensor network, ensuring that raw data never leaves the enterprise's secure perimeter.



```mermaid

graph TD

    subgraph "Enterprise Secure Environment"

        User((User)) -- "Asks a question" --> RAGApp["RAG Application"]

        RAGApp -- "Sends query to Validator" --> Validator

        Validator -- "Returns Document IDs" --> RAGApp

        RAGApp -- "Retrieves documents by ID" --> SecureDB[("Secure Datastore")]

        SecureDB -- "Document content" --> RAGApp

        RAGApp -- "Generates final response" --> User

        RawDocs[("Raw Documents")] --> EmbedGen["Embedding Generation"]

        EmbedGen -- "Indexing:
Embeddings + Doc IDs" --> Miner

    end



    subgraph "Bittensor Network (CERS Subnet)"

        Validator["Validator"] -- "Sends query embedding" --> Miner["Miner
(Vector Index)"]

        Miner -- "Returns relevant Document IDs" --> Validator

    end



    style SecureDB fill:#f9f,stroke:#333,stroke-width:2px

    style RawDocs fill:#f9f,stroke:#333,stroke-width:2px

```



Here’s how it works:

1.  **Data Stays On-Premise**: The enterprise's raw documents (the knowledge base) remain within their secure environment.

2.  **Embeddings, Not Data**: The enterprise generates numerical representations (embeddings) of their documents. Only these non-reversible embeddings, along with anonymous document IDs, are sent to the miners for indexing.

3.  **Miner's Role**: Miners store and maintain a searchable index of these embeddings. They perform similarity searches based on query embeddings provided by validators. **Miners never see the original document content.**

4.  **Secure Retrieval**: When a miner finds relevant results, it returns only the corresponding document IDs to the validator.

5.  **Final-Mile RAG**: The enterprise uses these IDs to retrieve the full document text from its own secure database to complete the RAG process.



This model ensures that companies can leverage the decentralized power of the Bittensor network for high-performance retrieval without ever compromising the confidentiality of their proprietary data.



---

## Getting Started



This guide provides the essential steps to get the Cohere Enterprise RAG Subnet up and running.



### Prerequisites



Before you begin, ensure you have the following installed:

- Python 3.8 or later

- `pip` and `venv`



You also need to have `bittensor` installed. If you haven't installed it yet, follow the instructions here.



### 1. Clone & Install



First, clone the repository and navigate into the directory:

```bash

git clone https://github.com/CohereAI-Labs/cers-subnet.git

cd cers-subnet

```



Next, set up a virtual environment and install the necessary dependencies. Using a virtual environment is highly recommended.

```bash

# Create and activate a virtual environment

python -m venv venv

source venv/bin/activate  # On Windows, use: venv\Scripts\activate



# Install the package in editable mode

pip install -e .

```



### 2. Configure Your Wallet



To interact with the Bittensor network, you need a wallet with a coldkey and a hotkey.

If you don't have one, create it using the Bittensor CLI (`btcli`):

```bash

# Create a new coldkey

btcli wallet new_coldkey --wallet.name my_wallet



# Create a new hotkey for your wallet

btcli wallet new_hotkey --wallet.name my_wallet --wallet.hotkey default

```



### 3. Run



You can participate in the subnet as either a miner or a validator.



#### For Miners



Miners are the core of the CERS subnet, responsible for indexing and retrieving information. For detailed instructions on how to set up and run a miner, including hardware requirements and configuration, please refer to the **[Miner README](./neurons/miners/README.md)**.



#### For Validators



Validators are responsible for evaluating miners and ensuring the integrity of the network. For detailed instructions on how to set up and run a validator, please refer to the **[Validator README](./neurons/validators/README.md)**.



For more detailed instructions, including how to register your keys and run on different networks (local, testnet, mainnet), please refer to our full documentation:

- Running Subnet Locally

- Running on the Test Network

- Running on the Main Network



## License

This repository is licensed under the MIT License. See the [LICENSE](LICENSE) file for details.