paint-brush
Using MinIO to Build a Retrieval Augmented Generation Chat Applicationby@minio
5,705 reads
5,705 reads

Using MinIO to Build a Retrieval Augmented Generation Chat Application

by MinIOSeptember 18th, 2024
Read on Terminal Reader
Read this story w/o Javascript

Too Long; Didn't Read

Building a production-grade RAG application demands a suitable data infrastructure to store, version, process, evaluate, and query chunks of data that comprise your proprietary corpus.
featured image - Using MinIO to Build a Retrieval Augmented Generation Chat Application
MinIO HackerNoon profile picture


It’s often been said that in the age of AI - data is your moat. To that end, building a production-grade RAG application demands a suitable data infrastructure to store, version, process, evaluate, and query chunks of data that comprise your proprietary corpus. Since MinIO takes a data-first approach to AI, our default initial infrastructure recommendation for a project of this type is to set up a Modern Data Lake (MinIO) and a vector database. While other ancillary tools may need to be plugged in along the way, these two infrastructure units are foundational. They will serve as the center of gravity for nearly all tasks subsequently encountered in getting your RAG application into production.


But you are in a conundrum. You’ve heard of these terms LLM and RAG before but beyond that you haven’t ventured much because of the unknown. But wouldn't it be nice if there was a “Hello World” or boilerplate app that can help you get started?


Don’t worry, I was in the same boat. So in this blog, we will demonstrate how to use MinIO to build a Retrieval Augmented Generation(RAG) based chat application using commodity hardware.


  • Use MinIO to store all the documents, processed chunks and the embeddings using the vector database.


  • Use MinIO's bucket notification feature to trigger events when adding or removing documents to a bucket


  • Webhook that consumes the event and process the documents using Langchain and saves the metadata and chunked documents to a metadata bucket


  • Trigger MinIO bucket notification events for newly added or removed chunked documents


  • A Webhook that consumes the events and generates embeddings and save it to the Vector Database (LanceDB) that is persisted in MinIO


Key Tools Used

  • MinIO - Object Store to persist all the Data
  • LanceDB - Serverless open-source Vector Database that persists data in object store
  • Ollama - To run LLM and embedding model locally (OpenAI API compatible)
  • Gradio - Interface through which to interact with RAG application
  • FastAPI - Server for the Webhooks that receives bucket notification from MinIO and exposes the Gradio App
  • LangChain & Unstructured - To Extract useful text from our documents and Chunk them for Embedding


Models Used

  • LLM - Phi-3-128K (3.8B Parameters)
  • Embeddings - Nomic Embed Text v1.5 (Matryoshka Embeddings/ 768 Dim, 8K context)

Start MinIO Server

You can download the binary if you don't have it already from here


# Run MinIO detached

!minio server ~/dev/data --console-address :9090 &


Start Ollama Server + Download LLM & Embedding Model

Download Ollama from here



# Start the Server

!ollama serve


# Download Phi-3 LLM

!ollama pull phi3:3.8b-mini-128k-instruct-q8_0


# Download Nomic Embed Text v1.5

!ollama pull nomic-embed-text:v1.5


# List All the Models

!ollama ls


Create A Basic Gradio App Using FastAPI to Test the Model

LLM_MODEL = "phi3:3.8b-mini-128k-instruct-q8_0"

EMBEDDING_MODEL = "nomic-embed-text:v1.5"

LLM_ENDPOINT = "http://localhost:11434/api/chat"

CHAT_API_PATH = "/chat"



def llm_chat(user_question, history):

    history = history or []

    user_message = f"**You**: {user_question}"

    llm_resp = requests.post(LLM_ENDPOINT,

                             json={"model": LLM_MODEL,

                                   "keep_alive": "48h", # Keep the model in-memory for 48 hours

                                   "messages": [

                                       {"role": "user",

                                        "content": user_question

                                        }

                                   ]},

                             stream=True)

    bot_response = "**AI:** "

    for resp in llm_resp.iter_lines():

        json_data = json.loads(resp)

        bot_response += json_data["message"]["content"]

        yield bot_response


import json

import gradio as gr

import requests

from fastapi import FastAPI, Request, BackgroundTasks

from pydantic import BaseModel

import uvicorn

import nest_asyncio


app = FastAPI()


with gr.Blocks(gr.themes.Soft()) as demo:

    gr.Markdown("## RAG with MinIO")

    ch_interface = gr.ChatInterface(llm_chat, undo_btn=None, clear_btn="Clear")

    ch_interface.chatbot.show_label = False

    ch_interface.chatbot.height = 600


demo.queue()

    


if __name__ == "__main__":

    nest_asyncio.apply()

    app = gr.mount_gradio_app(app, demo, path=CHAT_API_PATH)

    uvicorn.run(app, host="0.0.0.0", port=8808)

Test Embedding Model

import numpy as np


EMBEDDING_ENDPOINT = "http://localhost:11434/api/embeddings"

EMBEDDINGS_DIM = 768


def get_embedding(text):

    resp = requests.post(EMBEDDING_ENDPOINT,

                         json={"model": EMBEDDING_MODEL,

                               "prompt": text})

    return np.array(resp.json()["embedding"][:EMBEDDINGS_DIM], dtype=np.float16)



## Test with sample text

get_embedding("What is MinIO?")


Ingestion Pipeline Overview

Create MinIO Buckets

Use mc command or do it from UI

  • custom-corpus - To store all the documents
  • warehouse - To store all the metadata, chunks and vector embeddings


!mc alias set 'myminio' 'http://localhost:9000' 'minioadmin' 'minioadmin'


!mc mb myminio/custom-corpus

!mc mb myminio/warehouse

Create Webhook that Consumes Bucket Notifications from custom-corpus bucket

import json

import gradio as gr

import requests

from fastapi import FastAPI, Request

from pydantic import BaseModel

import uvicorn

import nest_asyncio


app = FastAPI()


@app.post("/api/v1/document/notification")

async def receive_webhook(request: Request):

    json_data = await request.json()

    print(json.dumps(json_data, indent=2))


with gr.Blocks(gr.themes.Soft()) as demo:

    gr.Markdown("## RAG with MinIO")

    ch_interface = gr.ChatInterface(llm_chat, undo_btn=None, clear_btn="Clear")

    ch_interface.chatbot.show_label = False


demo.queue()

    


if __name__ == "__main__":

    nest_asyncio.apply()

    app = gr.mount_gradio_app(app, demo, path=CHAT_API_PATH)

    uvicorn.run(app, host="0.0.0.0", port=8808)


## Test with sample text

get_embedding("What is MinIO?")


Create Webhook Event

In Console go to Events-> Add Event Destination -> Webhook


Fill the fields with Following values and hit save


Identifier - doc-webhook


Endpoint - http://localhost:8808/api/v1/document/notification


Click Restart MinIO at the top when pormpted to


(Note: You can also use mc for this)

In console go to Buckets (Administrator) -> custom-corpus -> Events


Fill the fields with Following values and hit save


ARN - Select the doc-webhook from dropdown


Select Events - Check PUT and DELETE


(Note: You can also use mc for this)


We have our first webhook setup

Now test by adding and removing an object

Extract data from the Documents and Chunk

We will use Langchain and Unstructured to read an object from MinIO and Split Documents in to multiples chunks


from langchain_text_splitters import RecursiveCharacterTextSplitter

from langchain_community.document_loaders import S3FileLoader


MINIO_ENDPOINT = "http://localhost:9000"

MINIO_ACCESS_KEY = "minioadmin"

MINIO_SECRET_KEY = "minioadmin"



# Split Text from a given document using chunk_size number of characters

text_splitter = RecursiveCharacterTextSplitter(chunk_size=1024,

                                               chunk_overlap=64,

                                               length_function=len)



def split_doc_by_chunks(bucket_name, object_key):

    loader = S3FileLoader(bucket_name,

                          object_key,

                          endpoint_url=MINIO_ENDPOINT,

                          aws_access_key_id=MINIO_ACCESS_KEY,

                          aws_secret_access_key=MINIO_SECRET_KEY)

    docs = loader.load()

    doc_splits = text_splitter.split_documents(docs)

    return doc_splits


# test the chunking

split_doc_by_chunks("custom-corpus", "The-Enterprise-Object-Store-Feature-Set.pdf")

Add the Chunking logic to Webhook

Add the chunk logic to webhook and save the metadata and chunks to warehouse bucket


import urllib.parse

import s3fs


METADATA_PREFIX = "metadata"


# Using s3fs to save and delete objects from MinIO

s3 = s3fs.S3FileSystem()



# Split the documents and save the metadata to warehouse bucket

def create_object_task(json_data):

    for record in json_data["Records"]:

        bucket_name = record["s3"]["bucket"]["name"]

        object_key = urllib.parse.unquote(record["s3"]["object"]["key"])

        print(record["s3"]["bucket"]["name"],

              record["s3"]["object"]["key"])


        doc_splits = split_doc_by_chunks(bucket_name, object_key)


        for i, chunk in enumerate(doc_splits):

            source = f"warehouse/{METADATA_PREFIX}/{bucket_name}/{object_key}/chunk_{i:05d}.json"

            with s3.open(source, "w") as f:

                f.write(chunk.json())

    return "Task completed!"



def delete_object_task(json_data):

    for record in json_data["Records"]:

        bucket_name = record["s3"]["bucket"]["name"]

        object_key = urllib.parse.unquote(record["s3"]["object"]["key"])

        s3.delete(f"warehouse/{METADATA_PREFIX}/{bucket_name}/{object_key}", recursive=True)

    return "Task completed!"

Update FastAPI server with the new logic

import json

import gradio as gr

import requests

from fastapi import FastAPI, Request, BackgroundTasks

from pydantic import BaseModel

import uvicorn

import nest_asyncio


app = FastAPI()


@app.post("/api/v1/document/notification")

async def receive_webhook(request: Request, background_tasks: BackgroundTasks):

    json_data = await request.json()

    if json_data["EventName"] == "s3:ObjectCreated:Put":

        print("New object created!")

        background_tasks.add_task(create_object_task, json_data)

    if json_data["EventName"] == "s3:ObjectRemoved:Delete":

        print("Object deleted!")

        background_tasks.add_task(delete_object_task, json_data)

    return {"status": "success"}


with gr.Blocks(gr.themes.Soft()) as demo:

    gr.Markdown("## RAG with MinIO")

    ch_interface = gr.ChatInterface(llm_chat, undo_btn=None, clear_btn="Clear")

    ch_interface.chatbot.show_label = False


demo.queue()

    


if __name__ == "__main__":

    nest_asyncio.apply()

    app = gr.mount_gradio_app(app, demo, path=CHAT_API_PATH)

    uvicorn.run(app, host="0.0.0.0", port=8808)

Add new webhook to process document metadata/chunks

Now that we have the first webhook working next step is the get all the chunks with metadata Generate the Embeddings and store it in the vector Database



import json

import gradio as gr

import requests

from fastapi import FastAPI, Request, BackgroundTasks

from pydantic import BaseModel

import uvicorn

import nest_asyncio


app = FastAPI()




@app.post("/api/v1/metadata/notification")

async def receive_metadata_webhook(request: Request, background_tasks: BackgroundTasks):

    json_data = await request.json()

    print(json.dumps(json_data, indent=2))


@app.post("/api/v1/document/notification")

async def receive_webhook(request: Request, background_tasks: BackgroundTasks):

    json_data = await request.json()

    if json_data["EventName"] == "s3:ObjectCreated:Put":

        print("New object created!")

        background_tasks.add_task(create_object_task, json_data)

    if json_data["EventName"] == "s3:ObjectRemoved:Delete":

        print("Object deleted!")

        background_tasks.add_task(delete_object_task, json_data)

    return {"status": "success"}


with gr.Blocks(gr.themes.Soft()) as demo:

    gr.Markdown("## RAG with MinIO")

    ch_interface = gr.ChatInterface(llm_chat, undo_btn=None, clear_btn="Clear")

    ch_interface.chatbot.show_label = False


demo.queue()

    


if __name__ == "__main__":

    nest_asyncio.apply()

    app = gr.mount_gradio_app(app, demo, path=CHAT_API_PATH)

    uvicorn.run(app, host="0.0.0.0", port=8808)


Creat Webhook Event

In Console go to Events-> Add Event Destination -> Webhook


Fill the fields with Following values and hit save


Identifier - metadata-webhook


Endpoint - http://localhost:8808/api/v1/metadata/notification


Click Restart MinIO at the top when prompted to


(Note: You can also use mc for this)

In console go to Buckets (Administrator) -> warehouse -> Events


Fill the fields with Following values and hit save


ARN - Select the metadata-webhook from dropdown


Prefix - metadata/


Suffix - .json


Select Events - Check PUT and DELETE


(Note: You can also use mc for this)


We have our first webhook setup

Now test by adding and removing an object in custom-corpus and see if this webhook gets triggered

Create LanceDB Vector Database in MinIO

Now that we have the basic webhook working, lets setup the lanceDB vector databse in MinIO warehouse bucket in which we will save all the embeddings and additional metadata fields


import os

import lancedb



# Set these environment variables for the lanceDB to connect to MinIO

os.environ["AWS_DEFAULT_REGION"] = "us-east-1"

os.environ["AWS_ACCESS_KEY_ID"] = MINIO_ACCESS_KEY

os.environ["AWS_SECRET_ACCESS_KEY"] = MINIO_SECRET_KEY

os.environ["AWS_ENDPOINT"] = MINIO_ENDPOINT

os.environ["ALLOW_HTTP"] = "True"



db = lancedb.connect("s3://warehouse/v-db/")


# list existing tables

db.table_names()


# Create a new table with pydantic schema

from lancedb.pydantic import LanceModel, Vector

import pyarrow as pa


DOCS_TABLE = "docs"

EMBEDDINGS_DIM = 768


table = None



class DocsModel(LanceModel):

    parent_source: str # Actual object/document source

    source: str # Chunk/Metadata source

    text: str # Chunked text

    vector: Vector(EMBEDDINGS_DIM, pa.float16()) # Vector to be stored



def get_or_create_table():

    global table

    if table is None and DOCS_TABLE not in list(db.table_names()):

        return db.create_table(DOCS_TABLE, schema=DocsModel)

    if table is None:

        table = db.open_table(DOCS_TABLE)

    return table


# Check if that worked

get_or_create_table()


# list existing tables

db.table_names()

Add Storing/removing data from lanceDB to metadata-webhook

import multiprocessing


EMBEDDING_DOCUMENT_PREFIX = "search_document"


# Add queue that keeps the processed meteadata in memory

add_data_queue = multiprocessing.Queue()

delete_data_queue = multiprocessing.Queue()


def create_metadata_task(json_data):

    for record in json_data["Records"]:

        bucket_name = record["s3"]["bucket"]["name"]

        object_key = urllib.parse.unquote(record["s3"]["object"]["key"])

        print(bucket_name,

              object_key)

        with s3.open(f"{bucket_name}/{object_key}", "r") as f:

            data = f.read()

            chunk_json = json.loads(data)

            embeddings = get_embedding(f"{EMBEDDING_DOCUMENT_PREFIX}: {chunk_json['page_content']}")

            add_data_queue.put({

                "text": chunk_json["page_content"],

                "parent_source": chunk_json.get("metadata", "").get("source", ""),

                "source": f"{bucket_name}/{object_key}",

                "vector": embeddings

            })

    return "Metadata Create Task Completed!"



def delete_metadata_task(json_data):

    for record in json_data["Records"]:

        bucket_name = record["s3"]["bucket"]["name"]

        object_key = urllib.parse.unquote(record["s3"]["object"]["key"])

        delete_data_queue.put(f"{bucket_name}/{object_key}")

    return "Metadata Delete Task completed!"

Add a scheduler that Processes Data from Queues

from apscheduler.schedulers.background import BackgroundScheduler

import pandas as pd


def add_vector_job():

    data = []

    table = get_or_create_table()


    while not add_data_queue.empty():

        item = add_data_queue.get()

        data.append(item)


    if len(data) > 0:

        df = pd.DataFrame(data)

        table.add(df)

        table.compact_files()

        print(len(table.to_pandas()))



def delete_vector_job():

    table = get_or_create_table()

    source_data = []

    while not delete_data_queue.empty():

        item = delete_data_queue.get()

        source_data.append(item)

    if len(source_data) > 0:

        filter_data = ", ".join([f'"{d}"' for d in source_data])

        table.delete(f'source IN ({filter_data})')

        table.compact_files()

        table.cleanup_old_versions()

        print(len(table.to_pandas()))



scheduler = BackgroundScheduler()


scheduler.add_job(add_vector_job, 'interval', seconds=10)

scheduler.add_job(delete_vector_job, 'interval', seconds=10)

Update FastAPI with the Vector Embedding Changes

import json

import gradio as gr

import requests

from fastapi import FastAPI, Request, BackgroundTasks

from pydantic import BaseModel

import uvicorn

import nest_asyncio


app = FastAPI()



@app.on_event("startup")

async def startup_event():

    get_or_create_table()

    if not scheduler.running:

        scheduler.start()



@app.on_event("shutdown")

async def shutdown_event():

    scheduler.shutdown()


@app.post("/api/v1/metadata/notification")

async def receive_metadata_webhook(request: Request, background_tasks: BackgroundTasks):

    json_data = await request.json()

    if json_data["EventName"] == "s3:ObjectCreated:Put":

        print("New Metadata created!")

        background_tasks.add_task(create_metadata_task, json_data)

    if json_data["EventName"] == "s3:ObjectRemoved:Delete":

        print("Metadata deleted!")

        background_tasks.add_task(delete_metadata_task, json_data)

    return {"status": "success"}


@app.post("/api/v1/document/notification")

async def receive_webhook(request: Request, background_tasks: BackgroundTasks):

    json_data = await request.json()

    if json_data["EventName"] == "s3:ObjectCreated:Put":

        print("New object created!")

        background_tasks.add_task(create_object_task, json_data)

    if json_data["EventName"] == "s3:ObjectRemoved:Delete":

        print("Object deleted!")

        background_tasks.add_task(delete_object_task, json_data)

    return {"status": "success"}


with gr.Blocks(gr.themes.Soft()) as demo:

    gr.Markdown("## RAG with MinIO")

    ch_interface = gr.ChatInterface(llm_chat, undo_btn=None, clear_btn="Clear")

    ch_interface.chatbot.show_label = False

    ch_interface.chatbot.height = 600


demo.queue()

    


if __name__ == "__main__":

    nest_asyncio.apply()

    app = gr.mount_gradio_app(app, demo, path=CHAT_API_PATH)

    uvicorn.run(app, host="0.0.0.0", port=8808)




Now that we have the Ingestion pipeline working let's integrate the final RAG pipeline.

Add Vector Search Capability

Now that we have the document ingested into the lanceDB let's add the search capability


EMBEDDING_QUERY_PREFIX = "search_query"


def search(query, limit=5):

    query_embedding = get_embedding(f"{EMBEDDING_QUERY_PREFIX}: {query}")

    res = get_or_create_table().search(query_embedding).metric("cosine").limit(limit)

    return res


# Lets test to see if it works


res = search("What is MinIO Enterprise Object Store Lite?")

res.to_list()

Prompt LLM to use the Relevant Documents

RAG_PROMPT = """

DOCUMENT:

{documents}


QUESTION:

{user_question}


INSTRUCTIONS:

Answer in detail the user's QUESTION using the DOCUMENT text above.

Keep your answer ground in the facts of the DOCUMENT. Do not use sentence like "The document states" citing the document.

If the DOCUMENT doesn't contain the facts to answer the QUESTION only Respond with "Sorry! I Don't know"

"""


context_df = []


def llm_chat(user_question, history):

    history = history or []

    global context_df

    # Search for relevant document chunks

    res = search(user_question)

    documents = " ".join([d["text"].strip() for d in res.to_list()]) 

    # Pass the chunks to LLM for grounded response

    llm_resp = requests.post(LLM_ENDPOINT,

                             json={"model": LLM_MODEL,

                                   "messages": [

                                       {"role": "user",

                                        "content": RAG_PROMPT.format(user_question=user_question, documents=documents)

                                        }

                                   ],

                                   "options": {

                                       # "temperature": 0,

                                       "top_p": 0.90,

                                   }},

                             stream=True)

    bot_response = "**AI:** "

    for resp in llm_resp.iter_lines():

        json_data = json.loads(resp)

        bot_response += json_data["message"]["content"]

        yield bot_response

    context_df = res.to_pandas()

    context_df = context_df.drop(columns=['source', 'vector'])



def clear_events():

    global context_df

    context_df = []

    return context_df

Update FastAPI Chat Endpoint to use RAG

import json

import gradio as gr

import requests

from fastapi import FastAPI, Request, BackgroundTasks

from pydantic import BaseModel

import uvicorn

import nest_asyncio


app = FastAPI()



@app.on_event("startup")

async def startup_event():

    get_or_create_table()

    if not scheduler.running:

        scheduler.start()



@app.on_event("shutdown")

async def shutdown_event():

    scheduler.shutdown()


@app.post("/api/v1/metadata/notification")

async def receive_metadata_webhook(request: Request, background_tasks: BackgroundTasks):

    json_data = await request.json()

    if json_data["EventName"] == "s3:ObjectCreated:Put":

        print("New Metadata created!")

        background_tasks.add_task(create_metadata_task, json_data)

    if json_data["EventName"] == "s3:ObjectRemoved:Delete":

        print("Metadata deleted!")

        background_tasks.add_task(delete_metadata_task, json_data)

    return {"status": "success"}


@app.post("/api/v1/document/notification")

async def receive_webhook(request: Request, background_tasks: BackgroundTasks):

    json_data = await request.json()

    if json_data["EventName"] == "s3:ObjectCreated:Put":

        print("New object created!")

        background_tasks.add_task(create_object_task, json_data)

    if json_data["EventName"] == "s3:ObjectRemoved:Delete":

        print("Object deleted!")

        background_tasks.add_task(delete_object_task, json_data)

    return {"status": "success"}


with gr.Blocks(gr.themes.Soft()) as demo:

    gr.Markdown("## RAG with MinIO")

    ch_interface = gr.ChatInterface(llm_chat, undo_btn=None, clear_btn="Clear")

    ch_interface.chatbot.show_label = False

    ch_interface.chatbot.height = 600

    gr.Markdown("### Context Supplied")

    context_dataframe = gr.DataFrame(headers=["parent_source", "text", "_distance"], wrap=True)

    ch_interface.clear_btn.click(clear_events, [], context_dataframe)


    @gr.on(ch_interface.output_components, inputs=[ch_interface.chatbot], outputs=[context_dataframe])

    def update_chat_context_df(text):

        global context_df

        if context_df is not None:

            return context_df

        return ""


demo.queue()

    


if __name__ == "__main__":

    nest_asyncio.apply()

    app = gr.mount_gradio_app(app, demo, path=CHAT_API_PATH)

    uvicorn.run(app, host="0.0.0.0", port=8808)


Were you able to go through and implement RAG based chat with MinIO as the data lake backend? We will in the near future do a webinar on this same topic where we will give you a live demo as we build this RAG based chat application.

RAGs-R-Us

As a developer focused on AI integration at MinIO, I am constantly exploring how our tools can be seamlessly integrated into modern AI architectures to enhance efficiency and scalability. In this article, we showed you how to integrate MinIO with Retrieval-Augmented Generation (RAG) to build a chat application. This is just the tip of the iceberg, to give you a boost in your quest to build more unique used cases for RAG and MinIO. Now you have the building blocks to do it. Let's do it!


If you have any questions on MinIO RAG integration be sure to reach out to us on Slack!