Generative AI Search (RAG) with Marqo

First, select your platform:

Marqo Cloud

Marqo Open Source

This guide will walk you through setting up your environment and using Marqo to perform generative AI (RAG) searches. Follow along for a step-by-step tutorial.

Full code:

If you have any questions or need help, visit our Community and ask in the get-help channel.

Step 1: Get Marqo Cloud API Key

First, we need to obtain our Marqo Cloud API Key. For more information on how you can obtain this, visit our article. Once you have obtained this, replace your_api_key with your actual API Key:

api_key = "your_api_key"

Step 2: Set up Base LLM

First, we need a base Large Language Model (LLM) to perform RAG searches. You can download the model we use in this example here. Once done so, place into a directory models in the same place your script is located.

Now, we set up the LLM:

from llama_cpp import Llama

# Initialize the Llama model
LLM = Llama(
    model_path="starter-guides/rag/models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
    n_ctx=4096,  # Increased context size for handling larger inputs
    n_gpu_layers=1,  # Enable GPU acceleration if available
)

# Generate the response using the Llama model
response_stream = LLM(
    "Who won gold in the women's 100 metre race at the Paris Olympics 2024?",
    max_tokens=512,  # Maximum number of tokens in the response
    stop=["Q:"],  # Stop specifies a list of stop sequences to end the generation
    stream=True,  # The response should be streamed
)

# Print the response
print("Just LLM Response:", first_paragraph)

We can obtain the response purely from the LLM before performing RAG.

This returns

Just LLM Response: Not Available

Clearly, the version of Llama 3.1 we're using is unable to tell us the answer to our query. Let's perform RAG and provide the LLM with some information about the Paris Olympics 2024 and specifically, articles that may help guide it to the correct answer.

Step 3: Use Marqo to Perform RAG

We can now create a Marqo client, add our documents to it and use this to improve the LLM's output.

from marqo import Client

# Set up the Client
mq = Client("https://api.marqo.ai", api_key=api_key)

# Create Marqo index
mq.create_index("news-index")

# Define documents
DOCUMENTS = [
    {
        "_id": "1",
        "date": "2024-08-03",
        "website": "www.bbc.com",
        "Title": "Alfred storms to Olympic 100m gold in Paris.",
        "Description": "Julien Alfred stormed to the women's 100m title at Paris 2024 to make history as St Lucia's first Olympic medallist. As the rain teemed down at a raucous Stade de France, Alfred, 23, dominated the final and sealed victory by a clear margin in a national record 10.72 seconds. American world champion Sha'Carri Richardson took silver in 10.87, with compatriot Melissa Jefferson (10.92) third. Great Britain's Daryll Neita finished four-hundredths of a second off the podium in fourth, crossing the line in 10.96. Neita produced the best finish by a British female athlete in an Olympic sprint final for 64 years but that will be of little consolation in her pursuit of a first individual global medal.",
    },
    # Add more documents here
]

# Indexing documents
mq.index("news-index").add_documents(DOCUMENTS, tensor_fields=["Title", "Description"])

# Peform search on Marqo index with the same question as before
results = mq.index("news-index").search(
    q="Who won gold in the women's 100 metre race at the Paris Olympics 2024?",
    filter_string=f"date:2024-08-03",
    limit=5,
)

# Print out the results
print(results)

# Use the results obtained by Marqo as part of context to the LLM when asking the question again
context = ""
for i, hit in enumerate(results["hits"]):
    title = hit["Title"]
    text = hit["Description"]
    context += f'Source {i}) {title} || {" ".join(text.split()[:60])}... \n'


def get_context_prompt(question, context):
    """LLM prompt with text-based context from Marqo search."""
    return f"Background: \n{context}\n\nQuestion: {question}\n\nAnswer:"


# Obtain the prompt with the content from Marqo
prompt_w_context = get_context_prompt(question=question, context=context)
print("Prompt to input into LLM: ", prompt_w_context)

# Generate the response
response_stream = LLM(prompt_w_context, max_tokens=512, stop=["Q:"], stream=True)

# Capture and print the output
response = "".join([item["choices"][0]["text"] for item in response_stream])
first_paragraph = response.split("\n\n", 1)[0].strip()
print("LLM & Marqo Response:", first_paragraph)

This returns:

LLM & Marqo Response: Julien Alfred from St Lucia.  She set a national record of 10.72 seconds.

We can clearly see that performing RAG has improved the LLM's ability to answer our queries.

Full Code

generative_search_cloud.py

#####################################################
### STEP 1. Define Marqo Cloud API Key
#####################################################

# To find your Marqo API Key, visit this article: https://marqo.ai/blog/finding-my-marqo-api-key
api_key = "your_api_key"

#####################################################
### STEP 2. Download LLM Model
#####################################################

# You can download the model directly from: https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/main/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf?download=true

# Once downloaded, place this into /models/

#####################################################
### STEP 3. Set Up LLM
#####################################################

from llama_cpp import Llama

# Initialize the Llama model
LLM = Llama(
    model_path="starter-guides/rag/models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
    n_ctx=4096,  # Increased context size for handling larger inputs
    n_gpu_layers=1  # Enable GPU acceleration if available
)

# Define the question to be asked
question = "Who won gold in the women's 100 metre race at the Paris Olympics 2024?"

# Generate the response using the Llama model
response_stream = LLM(
    question, 
    max_tokens=512,  # Maximum number of tokens in the response
    stop=["Q:"],  # Stop specifies a list of stop sequences to end the generation
    stream=True  # The response should be streamed
)

# Capture the output - the response_stream provides a stream of token choices; we join them to form a complete response
response = ''.join([item['choices'][0]['text'] for item in response_stream])

# Extract only the first paragraph from the response - Splits the response by newline and takes the first part
first_paragraph = response.split('\n', 1)[0].strip()

# Print the first paragraph of the response
print("Just LLM Response:", first_paragraph)

#####################################################
### STEP 4. Define Documents to Perform RAG
#####################################################

# Define documents you want to use as part of your RAG
DOCUMENTS = [
    {
        '_id': '1',
        'date': '2024-08-03',
        'website': 'www.bbc.com',
        'Title': "Alfred storms to Olympic 100m gold in Paris.",
        'Description': "Julien Alfred stormed to the women's 100m title at Paris 2024 to make history as St Lucia's first Olympic medallist. As the rain teemed down at a raucous Stade de France, Alfred, 23, dominated the final and sealed victory by a clear margin in a national record 10.72 seconds. American world champion Sha'Carri Richardson took silver in 10.87, with compatriot Melissa Jefferson (10.92) third. Great Britain's Daryll Neita finished four-hundredths of a second off the podium in fourth, crossing the line in 10.96. Neita produced the best finish by a British female athlete in an Olympic sprint final for 64 years but that will be of little consolation in her pursuit of a first individual global medal."
    },
    # Add more documents here
]

#####################################################
### STEP 5. Use Marqo to Perform RAG
#####################################################

from marqo import Client

# Your index name
index_name = 'news-index-open-source'

# Set up the Client
mq = Client(
    "https://api.marqo.ai", 
    api_key=api_key
)

# We create the index. Note if it already exists an error will occur
# as you cannot overwrite an existing index. For this reason, we delete
# any existing index 
try:
    mq.index(index_name).delete()
except:
    pass

# Create Marqo index
mq.create_index(index_name)

# Indexing documents 
mq.index(index_name).add_documents(DOCUMENTS, tensor_fields= ["Title", "Description"])

# Ensure only documents with this date are included
date = '2024-08-03'

# Peform search on Marqo index with the same question as before
results = mq.index(index_name).search(
    q=question, 
    filter_string=f"date:{date}", 
    limit=5)

# Print out the results
print(results)

# Use the results obtained by Marqo as part of context to the LLM when asking the question again
context = ''
for i, hit in enumerate(results['hits']):
    title =  hit['Title']
    text = hit['Description']
    context += f'Source {i}) {title} || {" ".join(text.split()[:60])}... \n'


def get_context_prompt(question, context):
    """ LLM prompt with text-based context from Marqo search. """
    return f'Background: \n{context}\n\nQuestion: {question}\n\nAnswer:'

# Obtain the prompt with the content from Marqo
prompt_w_context = get_context_prompt(question=question, context=context)
print("Prompt to input into LLM: ", prompt_w_context)

# Generate the response
response_stream = LLM(prompt_w_context, max_tokens=512, stop=["Q:"], stream=True)

# Capture and print the output
response = ''.join([item['choices'][0]['text'] for item in response_stream])
first_paragraph = response.split('\n\n', 1)[0].strip()
print("LLM & Marqo Response:", first_paragraph)

This guide will walk you through setting up your environment and using Marqo to perform generative AI (RAG) searches. Follow along for a step-by-step tutorial.

Full code: Generative AI Search (RAG) with Marqo Open Source Code

Step 1: Run Marqo

We need to get Marqo up and running. You can do this by executing the following command in your terminal:

docker pull marqoai/marqo:latest
docker rm -f marqo
docker run --name marqo -it -p 8882:8882 marqoai/marqo:latest

For more detailed instructions, check the Installation Guide.