Retrieval-Augmented Generation: Question Answering based on Custom Dataset with Open-sourced LangChain Library

---

This notebook’s CI test result for us-west-2 is as follows. CI test results in other regions can be found at the end of the notebook.

---

Many use cases such as building a chatbot require text (text2text) generation models like BloomZ 7B1, Flan T5 XXL, and Flan T5 UL2 to respond to user questions with insightful answers. The BloomZ 7B1, Flan T5 XXL, and Flan T5 UL2 models have picked up a lot of general knowledge in training, but we often need to ingest and use a large library of more specific information.

In this notebook we will demonstrate how to use BloomZ 7B1, Flan T5 XXL, and Flan T5 UL2 to answer questions using a library of documents as a reference, by using document embeddings and retrieval. The embeddings are generated from GPT-J-6B embedding model.

This notebook serves a template such that you can easily replace the example dataset by your own to build a custom question and asnwering application.

Step 1. Deploy large language model (LLM) and embedding model in SageMaker JumpStart

To better illustrate the idea, let’s first deploy all the models that are required to perform the demo. You can choose either deploying all three Flan T5 XL, BloomZ 7B1, and Flan UL2 models as the large language model (LLM) to compare their model performances, or select subset of the models based on your preference. To do that, you need modify the _MODEL_CONFIG_ python dictionary defined as below.

!pip install --upgrade sagemaker --quiet
!pip install ipywidgets==7.0.0 --quiet
!pip install langchain==0.0.148 --quiet
!pip install faiss-cpu --quiet

import time
import sagemaker, boto3, json
from sagemaker.session import Session
from sagemaker.model import Model
from sagemaker import image_uris, model_uris, script_uris, hyperparameters
from sagemaker.predictor import Predictor
from sagemaker.utils import name_from_base
from typing import Any, Dict, List, Optional
from langchain.embeddings import SagemakerEndpointEmbeddings
from langchain.llms.sagemaker_endpoint import ContentHandlerBase

sagemaker_session = Session()
aws_role = sagemaker_session.get_caller_identity_arn()
aws_region = boto3.Session().region_name
sess = sagemaker.Session()
model_version = "1.*"

def query_endpoint_with_json_payload(encoded_json, endpoint_name, content_type="application/json"):
    client = boto3.client("runtime.sagemaker")
    response = client.invoke_endpoint(
        EndpointName=endpoint_name, ContentType=content_type, Body=encoded_json
    )
    return response


def parse_response_model_flan_t5(query_response):
    model_predictions = json.loads(query_response["Body"].read())
    generated_text = model_predictions["generated_texts"]
    return generated_text


def parse_response_multiple_texts_bloomz(query_response):
    generated_text = []
    model_predictions = json.loads(query_response["Body"].read())
    for x in model_predictions[0]:
        generated_text.append(x["generated_text"])
    return generated_text

Deploy SageMaker endpoint(s) for large language models and GPT-J 6B embedding model. Please uncomment the entries as below if you want to deploy multiple LLM models to compare their performance.

_MODEL_CONFIG_ = {
    "huggingface-text2text-flan-t5-xxl": {
        "instance type": "ml.g5.12xlarge",
        "env": {"SAGEMAKER_MODEL_SERVER_WORKERS": "1", "TS_DEFAULT_WORKERS_PER_MODEL": "1"},
        "parse_function": parse_response_model_flan_t5,
        "prompt": """Answer based on context:\n\n{context}\n\n{question}""",
    },
    "huggingface-textembedding-gpt-j-6b": {
        "instance type": "ml.g5.24xlarge",
        "env": {"SAGEMAKER_MODEL_SERVER_WORKERS": "1", "TS_DEFAULT_WORKERS_PER_MODEL": "1"},
    },
    # "huggingface-textgeneration1-bloomz-7b1-fp16": {
    #     "instance type": "ml.g5.12xlarge",
    #     "env": {},
    #     "parse_function": parse_response_multiple_texts_bloomz,
    #     "prompt": """question: \"{question}"\\n\nContext: \"{context}"\\n\nAnswer:""",
    # },
    # "huggingface-text2text-flan-ul2-bf16": {
    #     "instance type": "ml.g5.24xlarge",
    #     "env": {
    #         "SAGEMAKER_MODEL_SERVER_WORKERS": "1",
    #         "TS_DEFAULT_WORKERS_PER_MODEL": "1"
    #     },
    #     "parse_function": parse_response_model_flan_t5,
    #     "prompt": """Answer based on context:\n\n{context}\n\n{question}""",
    # },
}

newline, bold, unbold = "\n", "\033[1m", "\033[0m"

for model_id in _MODEL_CONFIG_:
    endpoint_name = name_from_base(f"jumpstart-example-raglc-{model_id}")
    inference_instance_type = _MODEL_CONFIG_[model_id]["instance type"]

    # Retrieve the inference container uri. This is the base HuggingFace container image for the default model above.
    deploy_image_uri = image_uris.retrieve(
        region=None,
        framework=None,  # automatically inferred from model_id
        image_scope="inference",
        model_id=model_id,
        model_version=model_version,
        instance_type=inference_instance_type,
    )
    # Retrieve the model uri.
    model_uri = model_uris.retrieve(
        model_id=model_id, model_version=model_version, model_scope="inference"
    )
    model_inference = Model(
        image_uri=deploy_image_uri,
        model_data=model_uri,
        role=aws_role,
        predictor_cls=Predictor,
        name=endpoint_name,
        env=_MODEL_CONFIG_[model_id]["env"],
    )
    model_predictor_inference = model_inference.deploy(
        initial_instance_count=1,
        instance_type=inference_instance_type,
        predictor_cls=Predictor,
        endpoint_name=endpoint_name,
    )
    print(f"{bold}Model {model_id} has been deployed successfully.{unbold}{newline}")
    _MODEL_CONFIG_[model_id]["endpoint_name"] = endpoint_name

Step 2. Ask a question to LLM without providing the context

To better illustrate why we need retrieval-augmented generation (RAG) based approach to solve the question and anwering problem. Let’s directly ask the model a question and see how they respond.

question = "Which instances can I use with Managed Spot Training in SageMaker?"

payload = {
    "text_inputs": question,
    "max_length": 100,
    "num_return_sequences": 1,
    "top_k": 50,
    "top_p": 0.95,
    "do_sample": True,
}

list_of_LLMs = list(_MODEL_CONFIG_.keys())
list_of_LLMs.remove("huggingface-textembedding-gpt-j-6b")  # remove the embedding model


for model_id in list_of_LLMs:
    endpoint_name = _MODEL_CONFIG_[model_id]["endpoint_name"]
    query_response = query_endpoint_with_json_payload(
        json.dumps(payload).encode("utf-8"), endpoint_name=endpoint_name
    )
    generated_texts = _MODEL_CONFIG_[model_id]["parse_function"](query_response)
    print(f"For model: {model_id}, the generated output is: {generated_texts[0]}\n")

You can see the generated answer is wrong or doesn’t make much sense.

Step 3. Improve the answer to the same question using prompt engineering with insightful context

To better answer the question well, we provide extra contextual information, combine it with a prompt, and send it to model together with the question. Below is an example.

context = """Managed Spot Training can be used with all instances supported in Amazon SageMaker. Managed Spot Training is supported in all AWS Regions where Amazon SageMaker is currently available."""

parameters = {
    "max_length": 200,
    "num_return_sequences": 1,
    "top_k": 250,
    "top_p": 0.95,
    "do_sample": False,
    "temperature": 1,
}

for model_id in list_of_LLMs:
    endpoint_name = _MODEL_CONFIG_[model_id]["endpoint_name"]

    prompt = _MODEL_CONFIG_[model_id]["prompt"]

    text_input = prompt.replace("{context}", context)
    text_input = text_input.replace("{question}", question)
    payload = {"text_inputs": text_input, **parameters}

    query_response = query_endpoint_with_json_payload(
        json.dumps(payload).encode("utf-8"), endpoint_name=endpoint_name
    )
    generated_texts = _MODEL_CONFIG_[model_id]["parse_function"](query_response)
    print(
        f"{bold}For model: {model_id}, the generated output is: {generated_texts[0]}{unbold}{newline}"
    )

The output from step 3 tells us the chance to get the correct response significantly correlates with the insightful context you send into the LLM.

Now, the question becomes where can I find the insightful context based on the user query? The answer is to use a pre-stored knowledge data base with retrieval augmented generation, as shown in step 4 below.

Step 4. Use RAG based approach with LangChain and SageMaker endpoints to build a simplified question and answering application.

We plan to use document embeddings to fetch the most relevant documents in our document knowledge library and combine them with the prompt that we provide to LLM.

To achieve that, we will do following.

1. Generate embedings for each of document in the knowledge library with SageMaker GPT-J-6B embedding model.

2. Identify top K most relevant documents based on user query.

   • 2.1 For a query of your interest, generate the embedding of the query using the same embedding model.

   • 2.2 Search the indexes of top K most relevant documents in the embedding space using in-memory Faiss search.

   • 2.3 Use the indexes to retrieve the corresponded documents.

3. Combine the retrieved documents with prompt and question and send them into SageMaker LLM.

Note: The retrieved document/text should be large enough to contain enough information to answer a question; but small enough to fit into the LLM prompt – maximum sequence length of 1024 tokens.

Wrap up our SageMaker endpoints for embedding model into langchain.embeddings.SagemakerEndpointEmbeddings. That requires a small overwritten of SagemakerEndpointEmbeddings class to make it compatible with SageMaker embedding mdoel.

from langchain.embeddings.sagemaker_endpoint import EmbeddingsContentHandler


class SagemakerEndpointEmbeddingsJumpStart(SagemakerEndpointEmbeddings):
    def embed_documents(self, texts: List[str], chunk_size: int = 5) -> List[List[float]]:
        """Compute doc embeddings using a SageMaker Inference Endpoint.

        Args:
            texts: The list of texts to embed.
            chunk_size: The chunk size defines how many input texts will
                be grouped together as request. If None, will use the
                chunk size specified by the class.

        Returns:
            List of embeddings, one for each text.
        """
        results = []
        _chunk_size = len(texts) if chunk_size > len(texts) else chunk_size

        for i in range(0, len(texts), _chunk_size):
            response = self._embedding_func(texts[i : i + _chunk_size])
            print
            results.extend(response)
        return results


class ContentHandler(EmbeddingsContentHandler):
    content_type = "application/json"
    accepts = "application/json"

    def transform_input(self, prompt: str, model_kwargs={}) -> bytes:
        input_str = json.dumps({"text_inputs": prompt, **model_kwargs})
        return input_str.encode("utf-8")

    def transform_output(self, output: bytes) -> str:
        response_json = json.loads(output.read().decode("utf-8"))
        embeddings = response_json["embedding"]
        return embeddings


content_handler = ContentHandler()

embeddings = SagemakerEndpointEmbeddingsJumpStart(
    endpoint_name=_MODEL_CONFIG_["huggingface-textembedding-gpt-j-6b"]["endpoint_name"],
    region_name=aws_region,
    content_handler=content_handler,
)

Next, we wrap up our SageMaker endpoints for LLM into langchain.llms.sagemaker_endpoint.SagemakerEndpoint.

from langchain.llms.sagemaker_endpoint import LLMContentHandler, SagemakerEndpoint

parameters = {
    "max_length": 200,
    "num_return_sequences": 1,
    "top_k": 250,
    "top_p": 0.95,
    "do_sample": False,
    "temperature": 1,
}


class ContentHandler(LLMContentHandler):
    content_type = "application/json"
    accepts = "application/json"

    def transform_input(self, prompt: str, model_kwargs={}) -> bytes:
        input_str = json.dumps({"text_inputs": prompt, **model_kwargs})
        return input_str.encode("utf-8")

    def transform_output(self, output: bytes) -> str:
        response_json = json.loads(output.read().decode("utf-8"))
        return response_json["generated_texts"][0]


content_handler = ContentHandler()

sm_llm = SagemakerEndpoint(
    endpoint_name=_MODEL_CONFIG_["huggingface-text2text-flan-t5-xxl"]["endpoint_name"],
    region_name=aws_region,
    model_kwargs=parameters,
    content_handler=content_handler,
)

Now, let’s download the example data and prepare it for demonstration. We will use Amazon SageMaker FAQs as knowledge library. The data are formatted in a CSV file with two columns Question and Answer. We use the Answer column as the documents of knowledge library, from which relevant documents are retrieved based on a query.

For your purpose, you can replace the example dataset of your own to build a custom question and answering application.

original_data = "s3://jumpstart-cache-prod-us-east-2/training-datasets/Amazon_SageMaker_FAQs/"

!mkdir -p rag_data
!aws s3 cp --recursive $original_data rag_data

For the case when you have data saved in multiple subsets. The following code will read all files that end with .csv and concatenate them together. Please ensure each csv file has the same format.

import glob
import os
import pandas as pd

all_files = glob.glob(os.path.join("rag_data/", "*.csv"))

df_knowledge = pd.concat(
    (pd.read_csv(f, header=None, names=["Question", "Answer"]) for f in all_files),
    axis=0,
    ignore_index=True,
)

Drop the Question column as it is not used in this demonstration.

df_knowledge.drop(["Question"], axis=1, inplace=True)

df_knowledge.head(5)

df_knowledge.to_csv("rag_data/processed_data.csv", header=False, index=False)

from langchain.chains import RetrievalQA
from langchain.llms import OpenAI
from langchain.document_loaders import TextLoader
from langchain.indexes import VectorstoreIndexCreator
from langchain.vectorstores import Chroma, AtlasDB, FAISS
from langchain.text_splitter import CharacterTextSplitter
from langchain import PromptTemplate
from langchain.chains.question_answering import load_qa_chain
from langchain.document_loaders.csv_loader import CSVLoader

Use langchain to read the csv data. There are multiple built-in functions in LangChain to read different format of files such as txt, html, and pdf. For details, see LangChain document loaders.

loader = CSVLoader(file_path="rag_data/processed_data.csv")

documents = loader.load()
# text_splitter = CharacterTextSplitter(chunk_size=300, chunk_overlap=0)
# texts = text_splitter.split_documents(documents) ### if you use langchain.document_loaders.TextLoader to load text file. You can uncomment the code
## to split the text.

Now, we can build an QA application. LangChain makes it extremly simple with following few lines of code.

Based on the question below, we can achieven the points in Step 4 with just a few lines of code as shown below.

question

index_creator = VectorstoreIndexCreator(
    vectorstore_cls=FAISS,
    embedding=embeddings,
    text_splitter=CharacterTextSplitter(chunk_size=300, chunk_overlap=0),
)

index = index_creator.from_loaders([loader])

index.query(question=question, llm=sm_llm)

Step 5. Customize the QA application above with different prompt.

Now, we see how simple it is to use LangChain to achieve question and answering application with just few lines of code. Let’s break down the above VectorstoreIndexCreator and see what’s happening under the hood. Furthermore, we will see how to incorporate a customize prompt rather than using a default prompt with VectorstoreIndexCreator.

Firstly, we generate embedings for each of document in the knowledge library with SageMaker GPT-J-6B embedding model.

docsearch = FAISS.from_documents(documents, embeddings)

question

Based on the question above, we then identify top K most relevant documents based on user query, where K = 3 in this setup.

docs = docsearch.similarity_search(question, k=3)

Print out the top 3 most relevant docuemnts as below.

docs

Finally, we combine the retrieved documents with prompt and question and send them into SageMaker LLM.

We define a customized prompt as below.

prompt_template = """Answer based on context:\n\n{context}\n\n{question}"""

PROMPT = PromptTemplate(template=prompt_template, input_variables=["context", "question"])

chain = load_qa_chain(llm=sm_llm, prompt=PROMPT)

Send the top 3 most relevant docuemnts and question into LLM to get a answer.

result = chain({"input_documents": docs, "question": question}, return_only_outputs=True)[
    "output_text"
]

Print the final answer from LLM as below, which is accurate.

result

Notebook CI Test Results

This notebook was tested in multiple regions. The test results are as follows, except for us-west-2 which is shown at the top of the notebook.