Skip to content

Latest commit

 

History

History
605 lines (449 loc) · 24.5 KB

README.md

File metadata and controls

605 lines (449 loc) · 24.5 KB

Large Language Models have revolutionized how we retrieve information or build search systems. Retrieval-augmented generation (RAG) methodology has become a common way to access or extract information.

This guide teaches you how to build a Retrieval-Augmented Generation application using SurrealDB, Fireworks, FastAPI, and Astro. By the end of this guide, you will be able to update the chatbot’s knowledge visually and obtain the latest and personalized responses to your queries.

Prerequisites

You'll need the following:

Tech Stack

The following technologies are used in creating our RAG application:

Technology Type Description
https://fastapi.tiangolo.com/ Framework A high performance framework to build APIs with Python 3.8+.
https://astro.build Framework Framework for building fast, modern websites with serverless backend support.
https://tailwindcss.com/ Framework CSS framework for building custom designs.
https://surrealdb.com/ Platform A multi-model database platform.
https://fireworks.ai/ Platform Lightning-fast Inference platform to run generative AI models.

High-Level Data Flow and Operations

This is a high-level architecture of how data is flowing and operations that take place 👇🏻

highlevel-diag

  • When a user asks a question, relevant vectors to the latest user question are queried from SurrealDB. Further, they are combined with the user messages to create a system context. The response is then streamed to the user from Fireworks hosted Llama 3.1 405B Instruct Model.
  • When a user updates the existing knowledge other system, vector embeddings with metadata are created for the particular information, and then pushed to SurrealDB
💡 You can find the code for the application in the [Github Repo](https://github.com/fw-ai/ai-starter-kits.git).

Step 1: Setup SurrealDB Server

You can find various methods to install and run the SurrealDB server in the documentation. Let's opt for installing SurrealDB using its dedicated install script for our scenario. In your terminal window, execute the following command:

curl --proto '=https' --tlsv1.2 -sSf https://install.surrealdb.com | sh

The above command attempts to install the latest version of SurrealDB (per your platform and CPU type) into the /usr/local/bin folder in your system.

Once that is done, execute the following command in your terminal window:

surreal start --log trace --user root --pass root --bind 0.0.0.0:4304 file:mydatabase.db

The above command does the following:

  • Starts the SurrealDB server at 0.0.0.0:4304 network address.
  • Enables trace level logging producing verbose logs in your terminal window.
  • Sets the user and password of the default database as root.
  • Creates the file mydatabase.db to persist data on your filesystem.

Step 2: Generate Fireworks AI API Key

Model inference requests to the Fireworks API require an API Key. To generate this API key, log in to your Fireworks account and navigate to API Keys. Enter a name for your API key and click the Create Key button to generate a new API key. Copy and securely store this token for later use as FIREWORKS_API_KEY environment variable.

Locally, set and export the FIREWORKS_API_KEY environment variable by executing the following command:

export FIREWORKS_API_KEY="<YOUR_FIREWORKS_API_KEY>"

Step 3: Create a new FastAPI application

First, let's start by creating a new project. You can create a new directory by executing the following command in your terminal window:

# Create and move to the new directory
mkdir chat-streaming
cd chat-streaming

Install Dependencies

Next, you can install the required dependencies by executing the following command in your terminal window:

pip install surrealdb
pip install fireworks-ai
pip install langchain langchain-community langchain_fireworks
pip install fastapi "uvicorn[standard]"

The above command installs the required libraries to run ASGI Server, FastAPI, Fireworks AI, SurrealDB and LangChain in your Python project.

Next, create a file main.py with the following code:

import uuid, os
from typing import List

# FastAPI
from fastapi import FastAPI
from pydantic import BaseModel

## Streaming Response utility
from fastapi.responses import StreamingResponse

## Enable CORS utility
from fastapi.middleware.cors import CORSMiddleware

# Fireworks SDK
import fireworks.client

# SurrealDB Vector Store SDK for LangChain
from langchain_community.vectorstores import SurrealDBStore

# Fireworks Embeddings Integration via LangChain
from langchain_fireworks import FireworksEmbeddings

The above code imports the following:

  • os module to use the environment variable you’ve set earlier.
  • List to denote a list of elements of specific type.
  • BaseModel class to define models of the request body FastAPI endpoints.
  • StreamingResponse class to generate streaming responses from FastAPI endpoints.
  • CORSMiddleware FastAPI middleware to enable Cross Origin Resource Sharing of FastAPI endpoints.
  • fireworks.client SDK for conveniently accessing Fireworks supported LLMs.
  • SurrealDBStore class by LangChain to use SurrealDB as vector store.
  • FireworksEmbeddings class via LangChain Fireworks integration to use Nomic AI Embeddings Model.

Define Data Models using Pydantic

To create the data types of request body in your FastAPI endpoints, append the following code in main.py file:

# Class representing the string of messages to be searched and embedded as system context.
class LearningMessages(BaseModel):
    messages: str

# Class representing a single message of the conversation between RAG application and user.
class Message(BaseModel):
    role: str
    content: str

# Class representing collection of messages above.
class Messages(BaseModel):
    messages: List[Message]

The above code defines three Pydantic models:

  • LearningMessages: a model that will store the input string with a single field called messages.
  • Message: a model that will store each message containing two fields, role and content.
  • Messages: a model that will store the input as a list of Message model.

Use Fireworks API Key

To set the Fireworks API key used by Fireworks AI module internally, append the following code in main.py file:

# Set the Fireworks API Key
fireworks.client.api_key = os.environ["FIREWORKS_API_KEY"]

The above code uses the os module to load the environment variable FIREWORKS_API_KEY as Firework’s API Key.

Use Fireworks Nomic AI Embeddings Model

To use FireworksEmbeddings class to create an embeddings generator using the nomic-ai/nomic-embed-text-v1.5, append the following code in main.py file:

# Load the nomic-embed-text-v1.5 embedding models via Langchain Fireworks Integration
embeddings = FireworksEmbeddings(model="nomic-ai/nomic-embed-text-v1.5",fireworks_api_key=os.getenv("FIREWORKS_API_KEY"))

Define SurrealDB Vector Store

To define the SurrealDB vector store configuration, append the following code in main.py file:

dburl = "ws://localhost:4304/rpc"
db_user = "root"
db_pass = "root"
vector_collection = "vectors"
vector_db = SurrealDBStore(dburl=dburl, db_user=db_user, db_pass=db_pass, collection=vector_collection, embedding_function=embeddings)

The above code uses the following values to establish a SurrealDB Vector Store with LangChain:

  • ws://localhost:4304/rpc as the database URL to establish a WebSocket connection with SurrealDB. Using a WebSocket connection allows to send and receive messages from SurrealDB using the WebSocket API.
  • root as both the username and password of the SurrealDB instance.
  • vectors as the collection name of the vector store to and from which the relevant vectors will be inserted and queried from.
  • Uses embeddings generator as the embedding function.

Initialize FastAPI App

To initialize a FastAPI application, append the following code in main.py file:

# Initialize FastAPI App
app = FastAPI()

# Add CORS middleware
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

The code above creates a FastAPI instance and uses the CORSMIddleware middleware to enable Cross Origin requests. This allows your frontend to successfully POST to the RAG application endpoints to fetch responses to the user query, regardless of the port it is running on.

Create a Knowledge Update API endpoint

To update application’s knowledge in realtime by generating vector embeddings and inserting them into SurrealDB, you’ll create an /update endpoint in your FastAPI application. Append the following code in main.py file:

@app.post('/update')
async def update(messages: LearningMessages):
    messages_json = messages.model_dump()["messages"].split(',')
    # Initialize SurrealDB
    await vector_db.initialize()
    # Create texts to be inserted into the Vector Store (Embeddings are generated automatically)
    metadatas = [{"len": len(t)} for t in messages_json]
    ids = [str(uuid.uuid4()) for _ in messages_json]
    await vector_db.aadd_texts(messages_json, metadatas=metadatas, ids=ids)

update(messages: LearningMessages) method -

  • Accepts a single string as messages containing comma (,) separated messages to be inserted in your SurrealDB vector store.
  • Awaits connection set up with SurrealDB.
  • Creates metadata list, each item being length of each message received as input.
  • Creates ids list, each item being a randomly generated id for each message received as input.
  • Using the embeddings generator passed as the embeddings function, it generates the vector embedding of each message. Alongwith each message’s metadata, it inserts the vector embedding into the SurrealDB vector store.

Create a Chat API endpoint

To generate personalized responses that uses the application’s existing knowledge, you’ll create an /chat endpoint in your FastAPI application. Append the following code in main.py file:

@app.post("/chat")
async def chat(messages: Messages):
    messages_json = (messages.model_dump())['messages']
    # Initialize SurrealDB
    await vector_db.initialize()
    # Create System Context
    knowledge = "Only answer what you know. If do not know, say it's an unknown. Following are the things you know of:\n"
    relevant_content = await vector_db.asimilarity_search(messages_json[-1]['content'])
    if relevant_content:
        for each_content in relevant_content:
            knowledge += each_content.page_content
    messages_json.insert(0, { "role": "system", "content": knowledge })
    # Create LLAMA Completion Responses
    response = fireworks.client.ChatCompletion.create(
        stream=True,
        prompt_or_messages=messages_json,
        model="accounts/fireworks/models/llama-v3p1-70b-instruct",
    )
    # Stream the response from requests.post
    return StreamingResponse(yield_content(response))

chat(messages: Messages) method -

  • Accepts a list of Message model as messages.
  • Awaits connection set up with SurrealDB.
  • Defines a system prompt to restrict it to answer what it already knows.
  • Performs a similarity search on the latest Message, which represents a user query.
  • Loops over all similar vector embeddings and appends them into the system prompt defined earlier.
  • Prepends a Message model, representing role of the system and it’s content as the system prompt created.
  • Uses fireworks Chat Completion API to stream LLAMA 3.1 70B Chat model context aware responses.
  • Returns a StreamingResponse using the yield_content function.

The yield_content function loops over each Document (received as the similar vector with it’s metadata), and streams the content value of it as part of the API response.

# Function to yield content from each choice
def yield_content(response):
    for chunk in response:
        if chunk.choices[0].delta.content:
            yield chunk.choices[0].delta.content

With all that done, here’s how our main.py will finally look like containing both the endpoints:

import uuid, os
from typing import List

# FastAPI
from fastapi import FastAPI
from pydantic import BaseModel
## Streaming Response utility
from fastapi.responses import StreamingResponse
## Enable CORS utility
from fastapi.middleware.cors import CORSMiddleware

# Fireworks SDK
import fireworks.client

# SurrealDB Vector Store SDK for LangChain
from langchain_community.vectorstores import SurrealDBStore

# Fireworks Embeddings Integration via LangChain
from langchain_fireworks import FireworksEmbeddings

# Class representing the string of messages to be searched and embedded as system context.
class LearningMessages(BaseModel):
    messages: str

# Class representing a single message of the conversation between RAG application and user.
class Message(BaseModel):
    role: str
    content: str

# Class representing collection of messages above.
class Messages(BaseModel):
    messages: List[Message]

# Set the Fireworks API Key
fireworks.client.api_key = os.getenv("FIREWORKS_API_KEY")

# Load the nomic-embed-text-v1.5 embedding models via Langchain Fireworks Integration
embeddings = FireworksEmbeddings(model="nomic-ai/nomic-embed-text-v1.5",fireworks_api_key=os.getenv("FIREWORKS_API_KEY"))

dburl = "ws://localhost:4304/rpc"
db_user = "root"
db_pass = "root"
vector_collection = "vectors"
vector_db = SurrealDBStore(dburl=dburl, db_user=db_user, db_pass=db_pass, collection=vector_collection, embedding_function=embeddings)

app = FastAPI()

# Add CORS middleware
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Function to yield content from each choice
def yield_content(response):
    for chunk in response:
        if chunk.choices[0].delta.content:
            yield chunk.choices[0].delta.content

@app.post('/update')
async def update(messages: LearningMessages):
    messages_json = messages.model_dump()["messages"].split(',')
    # Initialize SurrealDB
    await vector_db.initialize()
    # Create texts to be inserted into the Vector Store (Embeddings are generated automatically)
    metadatas = [{"len": len(t)} for t in messages_json]
    ids = [str(uuid.uuid4()) for _ in messages_json]
    await vector_db.aadd_texts(messages_json, metadatas=metadatas, ids=ids)
    
@app.post("/chat")
async def chat(messages: Messages):
    messages_json = (messages.model_dump())['messages']
    # Initialize SurrealDB
    await vector_db.initialize()
    # Create System Context
    knowledge = "Only answer what you know. If do not know, say it's an unknown. Following are the things you know of:\n"
    relevant_content = await vector_db.asimilarity_search(messages_json[-1]['content'])
    if relevant_content:
        for each_content in relevant_content:
            knowledge += each_content.page_content
    messages_json.insert(0, { "role": "system", "content": knowledge })
    # Create LLAMA Completion Responses
    response = fireworks.client.ChatCompletion.create(
        stream=True,
        prompt_or_messages=messages_json,
        model="accounts/fireworks/models/llama-v3p1-405b-instruct",
    )
    # Stream the response from requests.post
    return StreamingResponse(yield_content(response))

Run FastAPI App Locally

Execute the following command in another terminal window:

uvicorn main:app --reload
💡 Use Python virtual environments, to avoid conflicts with other packages. Simply run `./venv/bin/uvicorn main:app --reload` to make a clear distinction between global and local environments.

The app should be running on localhost:8000. Let’s keep it running while we create an user interface to invoke the endpoints to create responses to user queries.

Create a new Astro application

Let’s get started by creating a new Astro project. Open your terminal and run the following command:

npm create astro@latest chat-ui

npm create astro is the recommended way to scaffold an Astro project quickly.

When prompted, choose the following:

  • Empty when prompted on how to start the new project.
  • Yes when prompted whether to write Typescript.
  • Strict when prompted how strict Typescript should be.
  • Yes when prompted to whether install dependencies.
  • Yes when prompted to whether initialize a git repository.

Once that’s done, you can move into the project directory and start the app:

cd chat-ui
npm run dev

The app should be running on localhost:4321. Let's close the development server as we move on to integrate TailwindCSS into the application.

Add Tailwind CSS to the application

For styling the app, you will be using Tailwind CSS. Install and set up Tailwind CSS at the root of our project's directory by running:

npx astro add tailwind

When prompted, choose:

  • Yes when prompted to install the Tailwind dependencies.
  • Yes when prompted to generate a minimal tailwind.config.mjs file.
  • Yes when prompted to make changes to Astro configuration file.

With choices as above, the command finishes integrating TailwindCSS into your Astro project. It installed the following dependency:

  • tailwindcss: TailwindCSS as a package to scan your project files to generate corresponding styles.
  • @astrojs/tailwind: The adapter that brings Tailwind's utility CSS classes to every .astro file and framework component in your project.

To create reactive interfaces quickly, let’s move onto integrating React in your application.

Integrate React in your Astro project

To prototype the reactive user interface quickly, you are gonna use React as the library with Astro. In your terminal window, execute the following command:

npx astro add react

npx allows us to execute npm packages binaries without having to first install it globally.

When prompted, choose the following:

  • Yes when prompted whether to install the React dependencies.
  • Yes when prompted whether to make changes to Astro configuration file.
  • Yes when prompted whether to make changes to tsconfig.json file.

To create conversation user interface easily, let’s move onto installing an AI SDK in your application.

Install an AI SDK and Axios

In your terminal window, run the command below to install the necessary library for building the conversation user interface:

npm install ai axios

The above command installs the following:

  • ai library to build AI-powered streaming text and chat UIs.
  • axios library to make HTTP requests.

Build Conversation User Interface

Inside src directory, create a Chat.jsx file with the following code:

// File: src/Chat.jsx

import { useChat } from 'ai/react'

export default function () {
  const { messages, handleSubmit, input, handleInputChange } = useChat({
    api: 'http://localhost:8000/chat',
  })
  return (
    <form className="mt-12 flex w-full max-w-[300px] flex-col" onSubmit={handleSubmit}>
      <input
        id="input"
        name="prompt"
        value={input}
        onChange={handleInputChange}
        placeholder="What's your next question?"
        className="mt-3 rounded border px-2 py-1 outline-none focus:border-black"
      />
      <button className="mt-3 max-w-max rounded border px-3 py-1 outline-none hover:bg-black hover:text-white" type="submit">
        Ask &rarr;
      </button>
      {messages.map((message, i) => (
        <div className="mt-3 border-t pt-3" key={i}>
          {message.content}
        </div>
      ))}
    </form>
  )
}

chat.jsx does the following:

  • Imports the useChat hook by ai SDK to manage the conversation between user and the application. It takes care of saving the entire conversation (on the client-side) and using them as the request body when it calls the user defined api endpoint to fetch the response from chatbot.
  • Exports a React component that returns a form containing an <input> element to allow users to enter their query. It then loops over all the messages in the entire conversation, including the latest response to the user query.

Now, let’s create a component that will allow the user to supply some strings to the application to take into consideration before it answers any of the user query.

Build User Interface to Update Application’s Knowledge

Inside src directory, create a Update.jsx  file with the following code:

// File: src/Update.jsx

import axios from 'axios'
import { useState } from 'react'

export default function () {
  const [messages, setMessages] = useState('')
  return (
    <form
      className="mt-12 flex w-full max-w-[300px] flex-col"
      onSubmit={(e) => {
        e.preventDefault()
        axios.post('http://localhost:8000/update', {
          messages,
        })
      }}
    >
      <textarea
        value={messages}
        id="learn_messages"
        name="learn_messages"
        onChange={(e) => setMessages(e.target.value)}
        placeholder="Things to learn [seperated by comma (,)]"
        className="mt-3 rounded border px-2 py-1 outline-none focus:border-black"
      />
      <button className="mt-3 max-w-max rounded border px-3 py-1 outline-none hover:bg-black hover:text-white" type="submit">
        Learn &rarr;
      </button>
    </form>
  )
}

Update.jsx -

  • Imports axios library and useState hook by React.
  • Exports a React component that returns a form containing an <textarea> element to allow users to enter multiple strings, wherein each string is represented between comma(s).
  • Upon form submission, it POSTs the messages as JSON to the http://localhost:8000/update endpoint.
  • To use the React components on the home page of your Astro application, make the following changes in src/pages/index.astro file:
---
+ import Chat from '../Chat'
+ import Update from '../Update'
---

<html lang="en">
  <head>
    <meta charset="utf-8" />
    <link rel="icon" type="image/svg+xml" href="/favicon.svg" />
    <meta name="viewport" content="width=device-width" />
    <meta name="generator" content={Astro.generator} />
    <title>Astro</title>
  </head>
  <body class="flex w-screen flex-col items-center">
-   <h1>Astro</h1>
+   <Update client:load />
+   <Chat client:load />
  </body>
</html>

The changes above being with importing both the Chat and Update React components. Then, it uses Astro's client:load directive to make sure that both the components are loaded and hydrated immediately on the page.

Run Astro Application Locally

Run your Astro application by executing the following command in another terminal window:

npm run build && npm run preview

The app should be running on localhost:4321.

Conclusion

Congratulations, you created a Retrieval-Augmented Generation application using SurrealDB and Fireworks AI. With SurrealDB’s vector store, you are able to insert and update vector embeddings on the fly over WebSockets, and perform similarity search to user queries using vector embeddings generated internally for you.

Further, using Fireworks AI, you are able to invoke Llama 3.1 70B Chat model with system context and generate personalized responses to user queries.