Use Amazon SageMaker Studio to construct a RAG query answering resolution with Llama 2, LangChain, and Pinecone for quick experimentation

Retrieval Augmented Technology (RAG) means that you can present a big language mannequin (LLM) with entry to knowledge from exterior data sources resembling repositories, databases, and APIs with out the necessity to fine-tune it. When utilizing generative AI for query answering, RAG permits LLMs to reply questions with essentially the most related, up-to-date data and optionally cite their knowledge sources for verification.

A typical RAG resolution for data retrieval from paperwork makes use of an embeddings mannequin to transform the information from the information sources to embeddings and shops these embeddings in a vector database. When a consumer asks a query, it searches the vector database and retrieves paperwork which can be most just like the consumer’s question. Subsequent, it combines the retrieved paperwork and the consumer’s question in an augmented immediate that’s despatched to the LLM for textual content technology. There are two fashions on this implementation: the embeddings mannequin and the LLM that generates the ultimate response.

On this submit, we display the way to use Amazon SageMaker Studio to construct a RAG query answering resolution.

Utilizing notebooks for RAG-based query answering

Implementing RAG sometimes entails experimenting with varied embedding fashions, vector databases, textual content technology fashions, and prompts, whereas additionally debugging your code till you obtain a practical prototype. Amazon SageMaker affords managed Jupyter notebooks geared up with GPU situations, enabling you to quickly experiment throughout this preliminary part with out spinning up extra infrastructure. There are two choices for utilizing notebooks in SageMaker. The primary possibility is quick launch notebooks accessible by SageMaker Studio. In SageMaker Studio, the built-in growth atmosphere (IDE) purpose-built for ML, you’ll be able to launch notebooks that run on totally different occasion varieties and with totally different configurations, collaborate with colleagues, and entry extra purpose-built options for machine studying (ML). The second possibility is utilizing a SageMaker pocket book occasion, which is a totally managed ML compute occasion operating the Jupyter Pocket book app.

On this submit, we current a RAG resolution that augments the mannequin’s data with extra knowledge from exterior data sources to supply extra correct responses particular to a customized area. We use a single SageMaker Studio pocket book operating on an ml.g5.2xlarge occasion (1 A10G GPU) and Llama 2 7b chat hf, the fine-tuned model of Llama 2 7b, which is optimized for dialog use circumstances from Hugging Face Hub. We use two AWS Media & Leisure Weblog posts because the pattern exterior knowledge, which we convert into embeddings with the BAAI/bge-small-en-v1.5 embeddings. We retailer the embeddings in Pinecone, a vector-based database that provides high-performance search and similarity matching. We additionally talk about the way to transition from experimenting within the pocket book to deploying your fashions to SageMaker endpoints for real-time inference once you full your prototyping. The identical method can be utilized with totally different fashions and vector databases.

Resolution overview

The next diagram illustrates the answer structure.

Implementing the answer consists of two high-level steps: creating the answer utilizing SageMaker Studio notebooks, and deploying the fashions for inference.

Develop the answer utilizing SageMaker Studio notebooks

Full the next steps to begin creating the answer:

1. Load the Llama-2 7b chat mannequin from Hugging Face Hub within the pocket book.
2. Create a PromptTemplate with LangChain and use it to create prompts in your use case.
3. For 1–2 instance prompts, add related static textual content from exterior paperwork as immediate context and assess if the standard of the responses improves.
4. Assuming that the standard improves, implement the RAG query answering workflow:
   • Collect the exterior paperwork that may assist the mannequin higher reply the questions in your use case.
   • Load the BGE embeddings mannequin and use it to generate embeddings of those paperwork.
   • Retailer these embeddings in a Pinecone index.
   • When a consumer asks a query, carry out a similarity search in Pinecone and add the content material from essentially the most related paperwork to the immediate’s context.

Deploy the fashions to SageMaker for inference at scale

While you hit your efficiency targets, you’ll be able to deploy the fashions to SageMaker for use by generative AI functions:

1. Deploy the Llama-2 7b chat mannequin to a SageMaker real-time endpoint.
2. Deploy the BAAI/bge-small-en-v1.5 embeddings mannequin to a SageMaker real-time endpoint.
3. Use the deployed fashions in your query answering generative AI functions.

Within the following sections, we stroll you thru the steps of implementing this resolution in SageMaker Studio notebooks.

Conditions

To comply with the steps on this submit, you must have an AWS account and an AWS Id and Entry Administration (IAM) position with permissions to create and entry the answer assets. If you’re new to AWS, see Create a standalone AWS account.

To make use of SageMaker Studio notebooks in your AWS account, you want a SageMaker area with a consumer profile that has permissions to launch the SageMaker Studio app. If you’re new to SageMaker Studio, the Fast Studio setup is the quickest strategy to get began. With a single click on, SageMaker provisions the SageMaker area with default presets, together with organising the consumer profile, IAM position, IAM authentication, and public web entry. The pocket book for this submit assumes an ml.g5.2xlarge occasion sort. To evaluation or enhance your quota, open the AWS Service Quotas console, select AWS Providers within the navigation pane, select Amazon SageMaker, and consult with the worth for Studio KernelGateway apps operating on ml.g5.2xlarge situations.

After confirming your quota restrict, you must full the dependencies to make use of Llama 2 7b chat.

Llama 2 7b chat is on the market beneath the Llama 2 license. To entry Llama 2 on Hugging Face, you must full just a few steps first:

1. Create a Hugging Face account should you don’t have one already.
2. Full the shape “Request entry to the following model of Llama” on the Meta web site.
3. Request entry to Llama 2 7b chat on Hugging Face.

After you may have been granted entry, you’ll be able to create a brand new entry token to entry fashions. To create an entry token, navigate to the Settings web page on the Hugging Face web site.

You must have an account with Pinecone to make use of it as a vector database. Pinecone is on the market on AWS through the AWS Market. The Pinecone web site additionally affords the choice to create a free account that comes with permissions to create a single index, which is ample for the needs of this submit. To retrieve your Pinecone keys, open the Pinecone console and select API Keys.

Arrange the pocket book and atmosphere

To comply with the code on this submit, open SageMaker Studio and clone the next GitHub repository. Subsequent, open the pocket book studio-local-gen-ai/rag/RAG-with-Llama-2-on-Studio.ipynb and select the PyTorch 2.0.0 Python 3.10 GPU Optimized picture, Python 3 kernel, and ml.g5.2xlarge because the occasion sort. If that is your first time utilizing SageMaker Studio notebooks, consult with Create or Open an Amazon SageMaker Studio Pocket book.

To arrange the event atmosphere, you must set up the required Python libraries, as demonstrated within the following code:

%%writefile necessities.txt
sagemaker>=2.175.0
transformers==4.33.0
speed up==0.21.0
datasets==2.13.0
langchain==0.0.297
pypdf>=3.16.3
pinecone-client
sentence_transformers
safetensors>=0.3.3

!pip set up -U -r necessities.txt

Load the pre-trained mannequin and tokenizer

After you may have imported the required libraries, you’ll be able to load the Llama-2 7b chat mannequin together with its corresponding tokenizers from Hugging Face. These loaded mannequin artifacts are saved within the native listing inside SageMaker Studio. This lets you swiftly reload them into reminiscence at any time when you must resume your work at a unique time.

import torch

from transformers import (
	AutoTokenizer,
	LlamaTokenizer,
	LlamaForCausalLM,
	GenerationConfig,
	AutoModelForCausalLM
)
import transformers

tg_model_id = "meta-llama/Llama-2-7b-chat-hf" #the mannequin id in Hugging Face
tg_model_path = f"./tg_model/{tg_model_id}" #the native listing the place the mannequin will probably be saved

tg_model = AutoModelForCausalLM.from_pretrained(tg_model_id, token=hf_access_token,do_sample=True, use_safetensors=True, device_map="auto", torch_dtype=torch.float16
tg_tokenizer = AutoTokenizer.from_pretrained(tg_model_id, token=hf_access_token)

tg_model.save_pretrained(save_directory=tg_model_path, from_pt=True)
tg_tokenizer.save_pretrained(save_directory=tg_model_path, from_pt=True)

Ask a query that requires up-to-date data

Now you can begin utilizing the mannequin and ask questions. Llama-2 chat fashions count on the immediate to stick to the next format:

<s>[INST] <<SYS>>
system_prompt
<<SYS>>
{{ user_message }} [/INST]

You should utilize the PromptTemplate from LangChain to create a recipe primarily based on the immediate format, in an effort to simply create prompts going ahead:

from langchain import PromptTemplate

template = """<s>[INST] <<SYS>>nYou are an assistant for question-answering duties. You're useful and pleasant. Use the next items of retrieved context to reply the question. If you do not know the reply, you simply say I do not know. Use three sentences most and hold the reply concise.
<<SYS>>n
{context}n
{query} [/INST]
"""
prompt_template = PromptTemplate( template=template, input_variables=['context','question'] )

Let’s ask the mannequin a query that wants latest data from 2023. You should utilize LangChain and particularly the LLMChain sort of chain and cross as parameters the LLM, the immediate template you created earlier, and the query:

query = "When can I go to the AWS M&E Buyer Expertise Middle in New York Metropolis?"

tg_tokenizer.add_special_tokens( {"pad_token": "[PAD]"} )
tg_tokenizer.padding_side = "left"

tg_pipe = transformers.pipeline(job='text-generation', mannequin=tg_model, tokenizer=tg_tokenizer, num_return_sequences=1, eos_token_id=tg_tokenizer.eos_token_id, pad_token_id=tg_tokenizer.eos_token_id, max_new_tokens=400, temperature=0.7)

from langchain.chains import LLMChain
from langchain.llms import HuggingFacePipeline

llm=HuggingFacePipeline(pipeline=tg_pipe, model_kwargs={'temperature':0.7})
llm_chain = LLMChain(llm=llm, immediate=prompt_template)
no_context_response = llm_chain.predict(context="", query=query)
print(no_context_response)

We get the next generated reply:

Thanks for reaching out! The AWS M&E Buyer Expertise Middle in New York Metropolis is at the moment closed for visits because of the COVID-19 pandemic. Nevertheless, you’ll be able to examine their official web site or social media accounts for any updates on when the middle will reopen. Within the meantime, you’ll be able to discover their digital excursions and assets accessible on-line.

Enhance the reply by including context to the immediate

The reply we generated just isn’t totally true. Let’s see if we are able to enhance it by offering some context. You’ll be able to add an extract from the submit AWS declares new M&E Buyer Expertise Middle in New York, which incorporates updates on the subject from 2023:

context = """Media and leisure (M&E) prospects proceed to face challenges in creating extra content material, extra rapidly, and distributing it to extra endpoints than ever earlier than of their quest to please viewers globally. Amazon Internet Providers (AWS), together with AWS Companions, have showcased the fast evolution of M&E options for years at business occasions just like the Nationwide Affiliation of Broadcasters (NAB) Present and the Worldwide Broadcast Conference (IBC). Till now, AWS for M&E expertise demonstrations had been accessible on this method just some weeks out of the yr. Prospects are extra engaged than ever earlier than; they need to have increased high quality conversations relating to consumer expertise and media tooling. These conversations are greatest supported by having an interconnected resolution structure for reference. Scheduling a go to of the M&E Buyer Expertise Middle will probably be accessible beginning November thirteenth, please ship an e mail to AWS-MediaEnt-CXC@amazon.com."""

Use the LLMChain once more and cross the previous textual content as context:

context_response = llm_chain.predict(context=context, query=query)
print(context_response)

The brand new response solutions the query with up-to-date data:

You’ll be able to go to the AWS M&E Buyer Expertise Middle in New York Metropolis ranging from November thirteenth. Please ship an e mail to AWS-MediaEnt-CXC@amazon.com to schedule a go to.

We’ve got confirmed that by including the proper context, the mannequin’s efficiency is improved. Now you’ll be able to focus your efforts on discovering and including the proper context for the query requested. In different phrases, implement RAG.

Implement RAG query answering with BGE embeddings and Pinecone

At this juncture, you could determine on the sources of knowledge to boost the mannequin’s data. These sources may very well be inner webpages or paperwork inside your group, or publicly accessible knowledge sources. For the needs of this submit and for the sake of simplicity, we have now chosen two AWS Weblog posts printed in 2023:

These posts are already accessible as PDF paperwork within the knowledge challenge listing in SageMaker Studio for fast entry. To divide the paperwork into manageable chunks, you’ll be able to make use of the RecursiveCharacterTextSplitter methodology from LangChain:

from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.document_loaders import PyPDFDirectoryLoader

loader = PyPDFDirectoryLoader("./knowledge/")

paperwork = loader.load()

text_splitter=RecursiveCharacterTextSplitter(
     chunk_size=1000,
     chunk_overlap=5
)
docs = text_splitter.split_documents(paperwork)

Subsequent, use the BGE embeddings mannequin bge-small-en created by the Beijing Academy of Synthetic Intelligence (BAAI) that’s accessible on Hugging Face to generate the embeddings of those chunks. Obtain and save the mannequin within the native listing in Studio. We use fp32 in order that it could possibly run on the occasion’s CPU.

em_model_name = "BAAI/bge-small-en"
em_model_path = f"./em-model"

from transformers import AutoModel
# Load mannequin from HuggingFace Hub
em_model = AutoModel.from_pretrained(em_model_name,torch_dtype=torch.float32)
em_tokenizer = AutoTokenizer.from_pretrained(em_model_name,gadget="cuda")

# save mannequin to disk
em_tokenizer.save_pretrained(save_directory=f"{em_model_path}/mannequin",from_pt=True)
em_model.save_pretrained(save_directory=f"{em_model_path}/mannequin",from_pt=True)
em_model.eval()

Use the next code to create an embedding_generator perform, which takes the doc chunks as enter and generates the embeddings utilizing the BGE mannequin:

# Tokenize sentences
def tokenize_text(_input, gadget):
    return em_tokenizer(
        [_input], 
        padding=True, 
        truncation=True, 
        return_tensors="pt"
    ).to(gadget)

# Run embedding job as a perform with mannequin and textual content sentences as enter
def embedding_generator(_input, normalize=True):
    # Compute token embeddings
    with torch.no_grad():
        embedded_output = em_model(
            **tokenize_text(
                _input, 
                em_model.gadget
            )
        )
        sentence_embeddings = embedded_output[0][:, 0]
        # normalize embeddings
        if normalize:
            sentence_embeddings = torch.nn.practical.normalize(
                sentence_embeddings, 
                p=2, 
                dim=1
            )
    
    return sentence_embeddings[0, :].tolist()
    
sample_sentence_embedding = embedding_generator(docs[0].page_content)
print(f"Embedding measurement of the doc --->", len(sample_sentence_embedding))

On this submit, we display a RAG workflow utilizing Pinecone, a managed, cloud-native vector database that additionally affords an API for similarity search. You’re free to rewrite the next code to make use of your most well-liked vector database.

We initialize a Pinecone python consumer and create a brand new vector search index utilizing the embedding mannequin’s output size. We use LangChain’s built-in Pinecone class to ingest the embeddings we created within the earlier step. It wants three parameters: the paperwork to ingest, the embeddings generator perform, and the identify of the Pinecone index.

import pinecone
pinecone.init(
    api_key = os.environ["PINECONE_API_KEY"],
    atmosphere = os.environ["PINECONE_ENV"]
)
#examine if index already exists, if not we create it
index_name = "rag-index"
if index_name not in pinecone.list_indexes():
    pinecone.create_index(
        identify=index_name,
        dimension=len(sample_sentence_embedding), ## 384 for bge-small-en 
        metric="cosine"
    )

#insert the embeddings
from langchain.vectorstores import Pinecone
vector_store = Pinecone.from_documents(
    docs,
    embedding_generator,
    index_name=index_name
)

With the Llama-2 7B chat mannequin loaded into reminiscence and the embeddings built-in into the Pinecone index, now you can mix these parts to boost Llama 2’s responses for our question-answering use case. To attain this, you’ll be able to make use of the LangChain RetrievalQA, which augments the preliminary immediate with essentially the most related paperwork from the vector retailer. By setting return_source_documents=True, you acquire visibility into the precise paperwork used to generate the reply as a part of the response, permitting you to confirm the accuracy of the reply.

from langchain.chains import RetrievalQA
import textwrap

#helper methodology to enhance the readability of the response
def print_response(llm_response):
    temp = [textwrap.fill(line, width=100) for line in llm_response['result'].break up('n')]
    response="n".be a part of(temp)
    print(f"{llm_response['query']}n n{response}'n n Supply Paperwork:")
    for supply in llm_response["source_documents"]:
        print(supply.metadata)

llm_qa_chain = RetrievalQA.from_chain_type(
    llm=llm, #the Llama-2 7b chat mannequin
    chain_type="stuff",
    retriever=vector_store.as_retriever(search_kwargs={"ok": 2}), # carry out similarity search in Pinecone
    return_source_documents=True, #present the paperwork that had been used to reply the query
    chain_type_kwargs={"immediate": prompt_template}
)
print_response(llm_qa_chain(query))

We get the next reply:

Q: When can I go to the AWS M&E Buyer Expertise Middle in New York Metropolis?

A: I’m glad to assist! In response to the context, the AWS M&E Buyer Expertise Middle in New York Metropolis will probably be accessible for visits beginning on November thirteenth. You’ll be able to ship an e mail to AWS-MediaEnt-CXC@amazon.com to schedule a go to.’

Supply Paperwork:

{‘web page’: 4.0, ‘supply’: ‘knowledge/AWS declares new M&E Buyer Expertise Middle in New York Metropolis _ AWS for M&E Weblog.pdf’}

{‘web page’: 2.0, ‘supply’: ‘knowledge/AWS declares new M&E Buyer Expertise Middle in New York Metropolis _ AWS for M&E Weblog.pdf’}

Let’s attempt a unique query:

question2=" What number of awards have AWS Media Providers gained in 2023?"
print_response(llm_qa_chain(question2))

We get the next reply:

Q: What number of awards have AWS Media Providers gained in 2023?

A: In response to the weblog submit, AWS Media Providers have gained 5 business awards in 2023.’

Supply Paperwork:

{‘web page’: 0.0, ‘supply’: ‘knowledge/AWS Media Providers awarded business accolades _ AWS for M&E Weblog.pdf’}

{‘web page’: 1.0, ‘supply’: ‘knowledge/AWS Media Providers awarded business accolades _ AWS for M&E Weblog.pdf’}

After you may have established a ample degree of confidence, you’ll be able to deploy the fashions to SageMaker endpoints for real-time inference. These endpoints are totally managed and supply assist for auto scaling.

SageMaker affords giant mannequin inference utilizing Massive Mannequin Inference containers (LMIs), which we are able to make the most of to deploy our fashions. These containers come geared up with pre-installed open supply libraries like DeepSpeed, facilitating the implementation of performance-enhancing strategies resembling tensor parallelism throughout inference. Moreover, they use DJLServing as a pre-built built-in mannequin server. DJLServing is a high-performance, common model-serving resolution that provides assist for dynamic batching and employee auto scaling, thereby growing throughput.

In our method, we use the SageMaker LMI with DJLServing and DeepSpeed Inference to deploy the Llama-2-chat 7b and BGE fashions to SageMaker endpoints operating on ml.g5.2xlarge situations, enabling real-time inference. If you wish to comply with these steps your self, consult with the accompanying pocket book for detailed directions.

You’ll require two ml.g5.2xlarge situations for deployment. To evaluation or enhance your quota, open the AWS Service Quotas console, select AWS Providers within the navigation pane, select Amazon SageMaker, and consult with the worth for ml.g5.2xlarge for endpoint utilization.

The next steps define the method of deploying customized fashions for the RAG workflow on a SageMaker endpoint:

• Deploy the Llama-2 7b chat mannequin to a SageMaker real-time endpoint operating on an ml.g5.2xlarge occasion for quick textual content technology.
• Deploy the BAAI/bge-small-en-v1.5 embeddings mannequin to a SageMaker real-time endpoint operating on an ml.g5.2xlarge occasion. Alternatively, you’ll be able to deploy your individual embedding mannequin.
• Ask a query and use the LangChain RetrievalQA to reinforce the immediate with essentially the most related paperwork from Pinecone, this time utilizing the mannequin deployed within the SageMaker real-time endpoint:

# convert your native LLM into SageMaker endpoint LLM
llm_sm_ep = SagemakerEndpoint(
    endpoint_name=tg_sm_model.endpoint_name, # <--- Your text-gen mannequin endpoint identify
    region_name=area,
    model_kwargs={
        "temperature": 0.05, 
        "max_new_tokens": 512
    },
    content_handler=content_handler,
)

llm_qa_smep_chain = RetrievalQA.from_chain_type(
    llm=llm_sm_ep,  # <--- This makes use of SageMaker Endpoint mannequin for inference
    chain_type="stuff",
    retriever=vector_store.as_retriever(search_kwargs={"ok": 2}),
    return_source_documents=True,
    chain_type_kwargs={"immediate": prompt_template}
)

• Use LangChain to confirm that the SageMaker endpoint with the embedding mannequin works as anticipated in order that it may be used for future doc ingestion:

response_model = smr_client.invoke_endpoint(
    EndpointName=em_sm_model.endpoint_name, <--- Your embedding mannequin endpoint identify
    Physique=json.dumps({
        "textual content": "This can be a pattern textual content"
    }),
    ContentType="software/json",
)

outputs = json.masses(response_model["Body"].learn().decode("utf8"))['outputs']

Clear up

Full the next steps to wash up your assets:

• When you may have completed working in your SageMaker Studio pocket book, ensure you shut down the ml.g5.2xlarge occasion to keep away from any costs by selecting the cease icon. You may also arrange lifecycle configuration scripts to mechanically shut down assets when they aren’t used.

• For those who deployed the fashions to SageMaker endpoints, run the next code on the finish of the pocket book to delete the endpoints:

#delete your textual content technology endpoint
sm_client.delete_endpoint(
     EndpointName=tg_sm_model.endpoint_name
)
# delete your textual content embedding endpoint
sm_client.delete_endpoint(
      EndpointName=em_sm_model.endpoint_name
)

• Lastly, run the next line to delete the Pinecone index:

pinecone.delete_index(index_name)

Conclusion

SageMaker notebooks present a simple strategy to kickstart your journey with Retrieval Augmented Technology. They will let you experiment interactively with varied fashions, configurations, and questions with out spinning up extra infrastructure. On this submit, we confirmed the way to improve the efficiency of Llama 2 7b chat in a query answering use case utilizing LangChain, the BGE embeddings mannequin, and Pinecone. To get began, launch SageMaker Studio and run the pocket book accessible within the following GitHub repo. Please share your ideas within the feedback part!

In regards to the authors

Anastasia Tzeveleka is a Machine Studying and AI Specialist Options Architect at AWS. She works with prospects in EMEA and helps them architect machine studying options at scale utilizing AWS companies. She has labored on tasks in several domains together with Pure Language Processing (NLP), MLOps and Low Code No Code instruments.

Pranav Murthy is an AI/ML Specialist Options Architect at AWS. He focuses on serving to prospects construct, practice, deploy and migrate machine studying (ML) workloads to SageMaker. He beforehand labored within the semiconductor business creating giant laptop imaginative and prescient (CV) and pure language processing (NLP) fashions to enhance semiconductor processes. In his free time, he enjoys enjoying chess and touring.