shiguredd Knowledge Base

Azure provides a variety of compute options ranging from raw VMs to serverless functions:

• Azure App Service: Managed hosting for web apps, RESTful APIs, and mobile back ends.
• Azure Batch: Managed service for large-scale parallel and high-performance computing (HPC).
• Azure Container Apps: Serverless environment for containerized apps built on Kubernetes.
• Azure Container Instances (ACI): Simple way to run a single container or group of containers without managing VMs.
• Azure Functions: Event-driven serverless functions triggered by various events.
• Azure Kubernetes Service (AKS): Fully managed Kubernetes for container orchestration.
• Azure Red Hat OpenShift: Managed OpenShift clusters for production Kubernetes workloads.
• Azure Service Fabric: Distributed systems platform for microservices across Azure or on-premises.

Apply these strategies to reduce operational costs across the architecture:

• Lambda: Right-size memory (start at 256MB), set aggressive timeouts (30-60s), and use Reserved Concurrency to prevent cost spikes.
• Bedrock: Implement request throttling and token budgets per user. Use Claude 3 Haiku for simple tasks and Sonnet for complex ones.
• S3: Use S3 Intelligent-Tiering and lifecycle policies (Standard → IA → Glacier). Compress stored text using gzip.
• OpenSearch: Use Serverless mode and batch embedding storage operations.
• DynamoDB: Use On-Demand pricing and TTL for automatic record deletion.

To ensure data isolation in a multi-tenant serverless PDF QA system, use the following strategies:

• Identity Management: Use Amazon Cognito User Pools to isolate users.
• Database Isolation: Include the tenant_id or user_id as part of the DynamoDB partition key.
• Search Isolation: Use a dedicated OpenSearch index per tenant or implement strict metadata filtering within a shared index.

Follow these security guidelines to protect data and access in a serverless PDF QA environment:

• Authentication: Use Amazon Cognito User Pools with Multi-Factor Authentication (MFA).
• Authorization: Apply the principle of least privilege using IAM roles.
• Encryption:
  • S3: Use SSE-S3 or SSE-KMS.
  • DynamoDB: Enable encryption at rest.
  • API Gateway: Enforce TLS 1.2+.
• Isolation: Implement tenant-based data separation.
• Auditing: Enable AWS CloudTrail to log all API calls.
• Network: Use a VPC for Lambda functions if required for compliance.

Byte Pair Encoding (BPE) is a subword tokenization algorithm that iteratively merges the most frequent pair of adjacent symbols (bytes or characters) in a text corpus to create a vocabulary of subword units. This approach balances word-level and character-level representations, allowing models to handle out-of-vocabulary words by breaking them into known sub-units.

The BPE process follows these steps:

1. Initialize vocabulary with individual characters and their frequencies.
2. Count the frequency of all adjacent symbol pairs.
3. Merge the most frequent pair into a new single symbol.
4. Add the new symbol to the vocabulary.
5. Repeat until the target vocabulary size is reached or no more merges are possible.

To create a standalone page, add a Markdown or React file to the src/pages directory. The file system structure determines the URL path:

• src/pages/index.js maps to the root URL (/)
• src/pages/foo.md maps to /foo
• src/pages/foo/bar.js maps to /foo/bar

Follow these steps to implement the Q&A flow:

1. Receive user question via API Gateway.
2. Use a Lambda function to generate an embedding for the question.
3. Perform a vector similarity search in OpenSearch to retrieve the top 5-10 relevant document chunks.
4. Construct a prompt combining the retrieved context and the user's question.
5. Invoke Amazon Bedrock (Claude 3.5) to generate the final answer.
6. Return the answer to the user including source citations.