app.py

import gradio as gr
import boto3
import json
import os
import numpy as np
import botocore
import time
from scipy.spatial.distance import cosine as cosine_similarity

theme = gr.themes.Base(text_size='sm')

# Retrieve AWS credentials from environment variables

AWS_ACCESS_KEY_ID = os.getenv('AWS_ACCESS_KEY_ID')
AWS_SECRET_ACCESS_KEY = os.getenv('AWS_SECRET_ACCESS_KEY')
AWS_REGION = os.getenv('REGION_NAME')
AWS_SESSION = os.getenv('AWS_SESSION')
BUCKET_NAME = os.getenv('BUCKET_NAME')
EXTRACTIONS_PATH = os.getenv('EXTRACTIONS_PATH')

# Create AWS Bedrock client using environment variables
def create_bedrock_client():

    return boto3.client(
        'bedrock-runtime',
        region_name=AWS_REGION,
        aws_access_key_id=AWS_ACCESS_KEY_ID,
        aws_secret_access_key=AWS_SECRET_ACCESS_KEY,
        aws_session_token=AWS_SESSION
    )

def create_s3_client():
    
    # Create an S3 client
    return boto3.client(
        's3',
        aws_access_key_id=AWS_ACCESS_KEY_ID,
        aws_secret_access_key=AWS_SECRET_ACCESS_KEY,
        aws_session_token=AWS_SESSION
    )

def read_json_from_s3():

    response = s3_client.get_object(Bucket=BUCKET_NAME, Key=EXTRACTIONS_PATH)
    file_content = response['Body'].read().decode('utf-8')
    json_content = json.loads(file_content)
    
    return json_content

def get_titan_embedding(bedrock, doc_name, text, attempt=0, cutoff=10000):
    """
    Retrieves a text embedding for a given document using the Amazon Titan Embedding model.

    This function sends the provided text to the Amazon Titan text embedding model
    and retrieves the resulting embedding. It handles retries for throttling exceptions
    and input size limitations by recursively calling itself with adjusted parameters.

    Parameters:
        doc_name (str): The name of the document, used for logging and error messages.
        text (str): The text content to be processed by the Titan embedding model.
        attempt (int): The current attempt number (used in recursive calls to handle retries). Defaults to 0.
        cutoff (int): The maximum number of words to include from the input text if a ValidationException occurs due to input size limits. Defaults to 5000.

    Returns:
        dict or None: The embedding response from the Titan model as a dictionary, or None if the operation fails or exceeds the retry limits.
    """
    
    retries = 5
    
    try:
        model_id = 'amazon.titan-embed-text-v1' 
        accept = 'application/json' 
        content_type = 'application/json'
        
        body = json.dumps({
            "inputText": text,
        })

        # Invoke model 
        response = bedrock.invoke_model(
            body=body, 
            modelId=model_id, 
            accept=accept, 
            contentType=content_type
        )
        
        # Print response
        response_body = json.loads(response['body'].read())


    # Handle a few common client exceptions
    except botocore.exceptions.ClientError as error: 
        if error.response['Error']['Code'] == 'ThrottlingException':
            if attempt + 1 == retries:
                return None
                
            delay = 2 ** (attempt + 1);               
            time.sleep(delay)
            return get_titan_embedding(doc_name, text, attempt=attempt + 1)
            
        elif error.response['Error']['Code'] == 'ValidationException':
            # get chunks of text length 20000 characters
            text_chunks = [text[i:i+cutoff] for i in range(0, len(text), cutoff)]
            embeddings = []
            for chunk in text_chunks:
                embeddings.append(get_titan_embedding(bedrock, doc_name, chunk))
            
            # return the average of the embeddinngs
            return np.mean(embeddings, axis=0)
        
        else:
            yield f"Unhandled Exception when processing {doc_name}! : {error.response['Error']['Code']}"
            return None

    # Catch-all for any other exceptions
    except Exception as error:
        yield f"Unhandled Exception when processing {doc_name}: {type(error).__name__}"
        return None

    return response_body.get('embedding')

def ask_ds(message, history):


    if len(message) == 0:
        return
        
    question = message

    # RAG
    question_embedding = get_titan_embedding(bedrock_client, 'question', question)

    similar_documents = []
    for file, data in extractions.items():
        similarity = cosine_similarity(question_embedding, np.array(data['embedding']))
        similar_documents.append((file, similarity))

    similar_documents.sort(key=lambda x: x[1], reverse=False)

    similar_content = ''
    for file, _ in similar_documents[:5]:
        similar_content += extractions[file]['content'] + '\n'


    # Invoke
    response = bedrock_client.invoke_model_with_response_stream(
        modelId="anthropic.claude-3-sonnet-20240229-v1:0",
        body=json.dumps(
            {
                "anthropic_version": "bedrock-2023-05-31",
                "max_tokens": 4096,
                "system": f"""You are a helpful, excited assistant that answers questions about certain provided documents. 
                <Task>
                Your task is to review the provided relevant information and answer the user's question to the best of your ability. 
                Try to use only the information in the document to answer. Refrain from saying things like 'According to the relevant information provided'.

                Format your output nicely with sentences that are not too long. You should prefer lists or bullet points when applicable.
                Begin by thanking the user for their question, and at the end of your answer, say "Thank you for using Ask Dane Street!"
                </Task>
                
                <Relevant Information>
                {similar_content}
                </Relevant Information>""",
                "messages": [
                        {
                        "role": "user",
                        "content": [
                            {
                                "type": "text",
                                "text": message
                            }
                        ]
                    }
                ],
            }
        ),
    )

    all_text = ''
    stream = response.get('body')
    if stream:
        for event in stream:
            chunk = event.get('chunk')
            if chunk and json.loads(chunk.get('bytes').decode()):
                # check if delta is present
                try:
                    this_text = json.loads(chunk.get('bytes').decode()).get('delta').get('text')
                    all_text += this_text
                    yield all_text  # Stream the text back to the UI
                except:
                    pass

    output = '\n\nCheck out the following documents for more information:\n'
    for file, _ in similar_documents[:5]:
        output += f"\n{file.replace('.txt', '.pdf')}"

    yield all_text + output


bedrock_client = create_bedrock_client()
s3_client = create_s3_client()
extractions = read_json_from_s3()

demo = gr.ChatInterface(fn=ask_ds, title="AskDS_HR", multimodal=False, chatbot=gr.Chatbot(value=[(None, "")],),theme=theme)
demo.launch()