Reading time: 5 minutes

This is an in-depth look at how Brian RAG works and how developers can build their own Knowledge Base and integrate it into their apps and/or Discord and Telegram bots through the Brian API.

✨We have a special promotion running that offers projects the opportunity to create their own KB and integrate it through our API (up to 3k calls a month for 6 months). ✨

Reach out to us on X if you are interested in knowing more!

Introduction

At BrianknowsAI, our mission is to leverage advanced AI technologies to revolutionize the way people interact with the web3 world. We want to make it easier for anyone to use and understand blockchain-based services so that there is no such strong learning curve that can compromise the journey of experienced builders and newcomers.

What is AskBrian?

AskBrian is our intelligent query system that utilizes a Retrieval Augmented Generation (RAG) framework to provide precise and relevant answers to your Web3-related questions. This feature is not just a search engine; it’s an advanced AI tool that understands your queries and delivers comprehensive responses by accessing a vast repository of manually curated Web3 documentation.

AskBrian is available completely FREE on the Brian App for any users and for devs through our API*.

*The free tier has 1k API call a month.

The AskBrian feature, both on the Brian APP and API, can point to four different knowledge bases:

• Public KB: an overview of the main protocols and networks in the EVM space.
• Swarm: an overview of the Swarm ecosystem. Brian has received a grant from the Swarm Foundation.
• LIDO: an overview of the LIDO protocol, DAO, and how Ethereum staking works. Brian has received a grant from the LIDO DAO.
• ETHRome hackathon: an overview of the hackathon rules, tracks, and prizes. Useful as an AI assistant for hackers.

Understanding RAG

RAG leverages the capabilities of modern Large Language Models (LLMs) such as ChatGPT, Llama-3, Mistral, etc.

These models possess great language understanding and text generation capabilities. Modern LLMs can be used for various tasks, such as summarization, question answering, or named entity recognition, and they also show emergent capabilities, meaning they can perform some tasks that are not trained for.

However, it is well-known these language models have some significant issues:

1. The data on which they have been trained is up to a certain date, so they are not updated with the latest news.
2. They suffer from hallucinations. Often generating false responses. This can be misleading and harmful.

The RAG framework can handles both this issues, and currently is a SOTA mechanism to improve knowledge systems based on LLMs.

What we do for AskBrian is manually collect a set of documents in various formats, such as web pages, YouTube videos, and text files, and create a knowledge base from them. For example, one of our knowledge bases could be entirely dedicated to all Web3 events in 2024. In this human-in-the-loop (HINL) pipeline, we leverage both our team's expertise and LLMs' capabilities.

Once the documents are collected, they are divided (split) into various chunks. Often, documents are very long (consider an entire book, for instance), so breaking them into chunks facilitates future data retrieval.

After creating all our chunks, we embed them using a custom fine-tuned LLM.

This process projects a piece of text into a highly dimensional vector space. The embedding preserves the text’s semantics, so pieces of text discussing similar topics will be situated close to each other in the vector space.

This method allows us to identify which documents cover similar subjects.

What do we do with the chunks and embeddings at this point?

Suppose we receive a query from a user, such as “What is EthRome?”.

We project this query into the vector space, just like our documents, allowing us to identify which documents are similar to the query. This helps us determine which documents might contain the answer to the question.

Imagine that near the query “What is EthRome?” we find three documents (a Notion page, a YouTube video, and a tweet).

We can now use the text from these documents to formulate a richer question for the LLM we are using. For instance, one could use llama3, and construct the following contextualized prompt:

Based on the following context: 
Doc1 / Notion → ethrome is an event….. 
Doc2 / Youtube → welcome to ethrome, this is an hackathon…. 
Doc3 / X (Twitter) → join Ethrome… R
eply to the use question: What is EthRome?"

Now, the LLM will have a much richer context to work with, enabling it to provide a more accurate answer.

Additionally, it will be able to provide metadata from the documents used, such as the title, URL, and other relevant details.

Creating Custom Collections

One of the standout features of AskBrian is the ability to create custom collections on specific topics. If you need tailored information on a niche area within Web3, our team is ready to assist. Simply contact us, and we will help you curate a collection that meets your specific requirements.

One of our goal is to provide in the future a web application, where users can upload their knowledge bases and use it directly on Brian.

ℹ️ Stay tuned

👉🏻 Follow us on Twitter: https://twitter.com/BrianknowsAI

👉🏻 Try the Brian Public Beta: https://www.brianknows.org/

👉🏻 Read the Brian doc: https://docs.brianknows.org/

#buildinpublic — EP03: the LI.FI integration

Reading time: 3 minutes

This is the third Brian #buildinpublic, a series of articles where we share the amazing journey while building Brian. In this article, we’ll go over how the LI.FI integration has been crucial to offer users a vast choice of protocols for swapping and bridging tokens on several chains.

About Brian

Brian is your web3 AI assistant for making transactions, deploying contracts, and searching information and data by prompt in plain English. All the features available on the Brian App can be integrated into any protocol or wallet through the Brian APIs.

Brian’s mission is to set a new way of consuming web3 through text-based interaction and challenge the classic UX/UI of web3 that is often either too complex or too uninformative.

Read the full doc here: https://docs.brianknows.org/

About LiFi Protocol

LI.FI is a multi-chain liquidity aggregation protocol that supports any-2-any swaps by aggregating bridges and DEX aggregators across +20 networks. The ideal solution for wallets and/or defi protocols that want to offer their users the possibility to bridge across most of the chains and swap using most of the DEXs.

Popular wallets like Metamask, Blockwallet, Edge Wallet, or defi protocols like Summer.fi and StakeDAO integrated the Li.Fi protocol APIs.

Read the full doc here: https://docs.li.fi/

Making transactions by prompt

Brian transactions feature allows a user to perform on-chain operations such as “Bridge 100 USDC from Polygon to Gnosis” or “Can you swap 1 eth to dai on Arbitrum?” in a non-custodial way (!).

⚠️ Note: Brian is in private beta and to try these operations you have to be whitelisted. You can access this version by filling out this form here.

Our end goal is to provide a single point of access to perform major web3 transactions via prompts, a growing trend after the AI explosion thanks to OpenAI. Brian wants to be among the projects leading all the Web3xAI intersections. Performing a transaction with a prompt breaks all those UX barriers that are an obstacle both for newbies and experienced builders.

How much time do we lose every day finding the right bridge or the AMM where the token we want to swap has liquidity?

If we want to reach mass adoption, we need to simplify a lot of the consumer apps UX and, after years of infra-tech development, there is now the possibility to offer users easier ways to interact with the chain, always in a non-custodial way.

In the long run, we want to implement all the main operations that can be done with a prompt, like transfer, swap, bridge, liquidity providing, mint of an nft, register of an ENS, and so on, the possibilities are endless!

We started with the most common user operations, which include swap, bridge, and transfer. While these are simple operations, the most common ones currently on the Web3, the most important implementation challenge for us was to offer users the ability to swap tokens on major AMMs and bridge them across major chains. This is when we discovered the LI.FI protocol!

Problem: supporting several bridges and AMMs

What are the common problems devs find when they have to integrate several AMMs and bridges?

• Time-consuming integration
  Implementing different AMMs and bridges is complex because some of them work differently. This makes the 1-to-1 integration very time-consuming in both the development and testing phases.
• Error-proneness
  Integrating multiple AMMs and bridges manually increases the chances of errors and potential financial losses for users due to development mistakes/bugs.
• Maintenance and updates
  The implementation, often done through SDKs, needs constant updates. Consequently, maintaining and updating implementations for multiple AMMs and bridges becomes challenging.

One possible solution is to implement an aggregator protocol that aggregates a set of AMMs offering just one point of access. However, using these Aggregators APIs often requires a private key and the payment of a monthly subscription. For bridges, the situation is even worse.

Solution: Transaction by prompt feat LI.FI

For all these reasons above, we decided to implement LIFI, an aggregator that aggregates more than 25 AMMs and 15 bridges across 20+ chains (full list here).

LIFI exposes an API for finding the most optimal routes for conducting cross-chain swaps and bridges. It also offers an easy-to-use Typescript SDK for generating the transaction calldata.

This is ideal for our prompting system because it allows users to describe in natural language the action of swapping/bridging by passing common input parameters that anyone would use to perform such an operation. When the user prompts an action (eg. “I want to swap 10 USDC for MATIC on Polygon”), our LLM extracts the parameters and feeds them to the LIFI SDK, extracts the best possible route, and returns it to the user, ready to be executed.

In addition, LIFI’s tools have output parameters that enable us to build an informative frontend for users. This is crucial as it allows us to visually show users what they are about to execute.

Furthermore, LIFI offers us maximum flexibility as many optional parameters can be passed under the hood for customizing the aggregator behavior. For example, we can choose which AMMs and bridges users prefer to use or not use, as well as set parameters like custom slippage.

Brian X LI.FI

The integration of the technology offered by LI.FI and the innovation in web3 intent extraction offered by Brian is a perfect match from which emerges a new and useful solution for all web3 users.

To achieve the goal of breaking down barriers to interact with web3, Brian needs to offer a multichain and cross-chain product on which users can easily exchange and bridge as many tokens as possible. LI.FI is the perfect solution for this need!

We’re sure this is just the beginning of a long-term relationship with a protocol that will be widely integrated into any wallet and dApp in the web3 ecosystem.

📣 The Brian Private Beta is finally here! All the whitelisted users can research web3 info, deploy smart contracts, and finally perform transactions on several mainnets just by prompting in English. If you are a dev who wants to integrate Brian into a protocol or you are just curious about how Brian works, the developer documentation is available now!

⚠️ Remember, Brian is a non-custodial solution and an early-stage project, so be careful when you perform transactions, the Brian team will not be responsible for any problems that occur while testing Brian user interface.

This Private Beta has the goal of receiving as much feedback as possible and improving in every possible way the product, so if you have any thoughts, suggestions, or business inquiries, we encourage you to reach out to us via email at info@brianknows.org. Your feedback will be instrumental in shaping the future of Brian within the web3 and AI landscape!

🔐 Not whitelisted? No problem!

Even if you find yourself outside the current whitelist, don’t FOMO! We believe everyone should have the chance to experience the future with Brian. You’re invited to fill out the form and you’ll gain access in the next few days.

We’re committed to ensuring as many enthusiasts as possible join us on this journey during the Private Beta, so don’t miss out: whitelist your address here and join us in this important phase, every single user, transaction, and feedback is of incredible value for us.

🤔 But wait, what is Brian, and what’s new with the Private Beta?

Brian is the first AI assistant to interact with web3. Brian enables users to do multiple actions, like performing transactions, learning how web3 works, and deploying smart contracts just by prompting in English.

Users can prompt their intentions, and Brian will build the corresponding action. Eg. “ I want to swap 100 usdc to eth” or “Explain how Uniswap v3 works”.

Since its founding, in June at the ETHPrague 2023 hackathon, Brian was available only on testnets. Read more about our story here:
https://medium.com/@BrianknowsAI/buildinpublic-ep-01-brian-three-months-after-ethprague-e7316b3b1053

With the launch of the Private Beta, we are happy to announce that all whitelisted users can now perform actions on the following supported mainnets: Base, Ethereum, Arbitrum, Gnosis, Optimism, Polygon, and Linea. Goerli testnet is also available for those who want to try Brian without real tokens.

At the time of this article, Brian is composed of 3 tools:

• Non-custodial Transaction Builder
• Smart Contract Builder
• Search Tool

Non-custodial Transaction Builder: Users can express their intentions in plain English, and Brian will construct the transaction accordingly. This tool enhances the accessibility and user-friendliness of Web3 interactions, ensuring a smooth experience for both newcomers and experienced builders. Brian operates as a non-custodial service, requiring users to approve each action via Metamask or their chosen wallet before executing the transaction.

These are some of the actions supported by Brian:

• Swap of ETH and ERC-20 tokens
  eg. I want to swap 10 USDC to xDai.
• Bridge of ETH and ERC-20 tokens
  eg. Bridge 10 xDAI to USDC from Gnosis to Arbitrum.
• Transfer of ETH and ERC-20 tokens
  eg. I want to transfer 1100 USDC to 0xqc… on Ethereum.
• Crosschain swap of ETH and ERC-20 tokens
  eg. Bridge 100 GHO to GRT from Gnosis to Arbitrum.
• More operations in one prompt (Be careful: always specify the amount!)
  eg. I want to bridge 1000 USDC to Polygon and then swap 50 USDC to wETH.
• Independent operations in one prompt
  eg. swap 100 usdc to usdt and swap 100 usdc to dai

Smart Contract Builder: Smart Contract Builder is a unique tool that interprets English prompts to generate Solidity code. For example, a user might request to “deploy an ERC-20 token” and Brian responds with the corresponding Solidity code.

Moreover, Brian lets new developers understand what the code does with the “Explain this” feature. Finally, developers can edit the code and deploy the contract directly from Brian’s UI with the specific buttons.

Search Tool: The Search Tool is a research tool that utilizes the technology of a vector database to answer users’ questions about web3. The team manually selects resources from the most important protocols in the ecosystem, and Brian generates answers based on the closest resources to the user’s prompt. At the end of the response, the top three resources from which it was derived are shown, providing proof of the authority of the resources and allowing users to delve deeper into the answer by reading directly from the official documentation of the protocol.

📖 Developer Documentation is here!

In the past weeks, the team has also been working on the launch of the Developer Documentation for Brian, outlining the process for developers to seamlessly integrate Brian into third-party protocols or applications, including wallets, AMM interfaces, and Telegram/Discord bots.

All the functionalities available on the Brian App can be smoothly incorporated into other protocols using the forthcoming Brian APIs. While these APIs are not yet available, we encourage interested developers to explore the documentation. For any additional questions or clarifications, please feel free to reach out to us via email at info@brianknows.org.

🗣️Call for feedback, interact with Brian team!

Throughout the Private Beta, the Brian team will be accessible for feedback and one-on-one interviews with builders interested in learning more about the project and engaging with the team.

Join us in this exclusive Private Beta experience and be among the first to witness Brian in action showcasing its full capabilities with real-world transactions.

Are you ready to immerse yourself in the future of AI-powered intent-based applications?

ℹ️ Stay tuned

🐦 Follow us on Twitter: https://twitter.com/BrianknowsAI

🌐 Try Brian: https://www.brianknows.org/

💻 Developer Doc: https://brian-27.gitbook.io/developer-doc/

🔐 Register for the private beta: https://app.deform.cc/form/7fa80086-c215-461b-bef1-15ead9aed17a/

Reading time: 2 minutes

This is the second Brian #buildinpublic, where we share some Insights on the amazing journey we’re living while building Brian. In this article, we’ll go over the experience at the ETHRome Hackathon, where we have rewarded two great projects with the Brian bounty, and the Polygon Builders House, a special 4-week program with other nine startups and Polygon mentors.

🤖 Private Beta is coming! Whitelist your address here: https://app.deform.cc/form/7fa80086-c215-461b-bef1-15ead9aed17a/

🏛️ ETHRome Hackathon

ETHRome was the first web3 hackathon ever made in Rome and it took place from the 6th to the 8th at Citta dell’Altra Economia, a beautiful underground location in the center of the city.

For the Brian team, this has been one of the best experiences ever because most of the founders were among the ETHRome organizers and the event has been a success with incredible feedback from both sponsors and hackers. Brian presence at the hackathon was undeniably very strong, the team offered a bounty of 1k dollars* for the best AI-Web3 integrations and organized a side event with experts working in both fields.

* Paid in Brian API credits

The Brian Bounty winners are:

• CarbonaraAI: CarbonarAI tackled the same challenges that Brian aims to address: improving user experience in the web3 ecosystem. To achieve this, CarbonarAI empowers users to access valuable data through intent-based search. The team has done an excellent job in leveraging AI and machine learning algorithms to create a user-friendly web interface for querying on-chain data. Kudos to the CarbonarAI team for their innovative approach!
• Quo-Roma: Quo-Roma has introduced a chat feature that responds to users’ questions using an LLM. This innovation greatly helps newcomers to the web3 space by providing quick answers and facilitating their learning journey. This approach is a significant step forward in making blockchain technology accessible and user-friendly. Great work by the Quo-Roma team!

Congratulations to both teams! They will be among the first to have the opportunity to test the Brian API during the private beta phase!

We also organized a side event with a live demo of Brian and a roundtable session with experts from both AI and web3.

💜 Polygon Builders House

The Brian team has been accepted to the Polygon Builders house in Lisbon, a 1-month accelerator with mentorships from Polygon Labs team members and some prizes at the end of the program. This is a great opportunity for sharing insights and feedback with other web3 founders and learning from the Polygon mentors.

🔐 Private Beta

The Beta version will be released in early November 2023 and will support the most important EVM mainnets including Ethereum, Polygon, Gnosis, Optimism, Arbitrum and Base. In addition, the APIs and technical documentation for integrating Brian into other protocols will be launched during this phase.

Why whitelist your address for the private beta:

• Exclusive possibility of trying Brian on several mainnets before anyone else
• 1-to-1 sessions (tg chats and voice calls) to hear your feedback and suggestions
• Help us shape the future of web3 UX!

If you want to join the Private Beta and be among the first who try Brian on Mainnet, you have to whitelist your address here:
https://app.deform.cc/form/7fa80086-c215-461b-bef1-15ead9aed17a/

ℹ️ Stay tuned

This is the second Brian #buildinpublic article. We want to share with users, builders, and researchers all our journey and the daily challenges that we are facing both on web3 and AI.

👉🏻 Follow us on Twitter: https://twitter.com/BrianknowsAI

👉🏻 Try the Brian Open Alpha: https://www.brianknows.org/

👉🏻 Whitelist your address for the private beta: https://app.deform.cc/form/7fa80086-c215-461b-bef1-15ead9aed17a/

#buildinpublic — EP 01: Brian, three months after ETHPrague

Reading time: 5 minutes

In the first Brian #buildinpublic episode our team wants to go over all the incredible things that happened during the last three months. Everything started at the ETHPrague Hackathon in June and from that moment we have been working non-stop on the product features and GTM strategies. The next phase is the Brian Beta that we are launching in November 2023.

🤖 What is Brian?

Brian is the first ever non-custodial AI assistant for performing transactions by prompting. Users can prompt their intentions, and Brian will build the corresponding transaction:

• I want to bridge 1 eth to usdc
• I want to transfer 10 usd to 0xabc.
• I want to swap 10 usdc to eth from Polygon to Ethereum

Brian leverages LLMs and other AI techniques (like few-shot learning) for the user intent extraction, then builds the transaction, and the user has to confirm and execute it.

Here is a live demo of a swap on Gnosis (…with a great UX trick!)

When users do not have complete knowledge of the involved protocols, Brian also provides research by prompting on a custom-built knowledge base of updated web3 resources, for example:

• Explain the uniswap v3 protocol as I am 5 years old
• What is The Graph?
• What are the differences between Swarm and Ipfs?
• How does the Gnosis Chain work?
• What is the erc721vote?

In comparison with ChatGPT, Brian has not only updated web3 resources but also can give users the exact resources it has used for the answer generation. It means after asking “What is the erc721vote?” it gives users a short answer and the resource (https://docs.openzeppelin.com/contracts/4.x/api/governance).

This research service can be used also as a Discord/Telegram bot or as a personal assistant to hackers during Hackathon. For the ETHRome hackathon, there is a special “ETHRome 2023 Knowledge Base” that the users can leverage to ask for info about schedule, location, or bounties.

The combination of the two features (transaction and research by prompting) can help a lot of new and experienced builders in their daily interaction and comprehension of the web3 world.

🏆 ETHPrague and AugmentHack Hackathon

Brian is receiving great attention from both AI and web3 famous projects. At the ETHPrague Hackathon, the project won 5 first prizes:

• Building on Gnosis Chain
• Best GHO Hack (AAVE)
• Deploy on Taiko
• Best use of Existing Subgraphs (The Graph)
• Top project (Gateway.fm)

During ETHCC week, Brian’s team participated in the AugmentHack hackathon (a competition more focused on AI) and won three other prizes:

• Recognition Prize (AugmentHack)
• Best use of Bacalhau & Lilypad Network
• 1 Prize for the best use of Hugging Face

Moreover, recently it has been granted a $10k bounty from the Swarm Foundation and the team has applied for other several grants from blockchains and protocols.

🌐 Open Alpha

The Open Alpha was launched in June and it allows users to test some specific operations on the Goerli testnet. For the research feature, anyone can ask Brian questions on a limited list of web3 topics.

The supported actions on Goerli include:

1. Balance of (native coin/ERC20)
2. Total supply of (ERC20)
3. Token transfers (native coin/ERC20)
4. Swapping assets

The LLM model we’re using has been meticulously fine-tuned to allow each action to be expressed in multiple ways, offering users flexibility and ease of interaction.

Right now each action can be expressed in 80+ different ways such as “Can you swap 10 USDC for ETH” or “Can you exchange 10 USDC for ETH?” or “Can you buy 10 USDC of ETH?” for the swap operation.

🔥 Beta Release

The Beta version will be released in early November 2023 (date TBD) and will support the most important EVM mainnets including Ethereum, Polygon, Gnosis, Optimism, and Arbitrum. In addition, the SDK for integrating Brian into other protocols will be launched during this phase.

More info about the Beta launch will come in the next weeks!

ℹ️ Stay tuned

This is just the first Brian #buildinpublic article. We want to share with users, builders, and researchers all our journey and the daily challenges that we are facing both on web3 and AI.

👉🏻 Follow us on Twitter: https://twitter.com/BrianknowsAI

👉🏻 Try the Brian Open Alpha: https://www.brianknows.org/