Simple Chat - Application workflows

Simple Chat - Application workflows

• Content Safety

• Add your data (RAG Ingestion)

• Return to Main

Content Safety - Workflow

1. User Sends Message: A user types a message in the chat interface.
2. Content Safety Interrogation (If Enabled):
   • Before the message reaches any backend service (AI model, Search, Image Gen, etc.), it is sent to the configured Azure AI Content Safety endpoint.
   • Content Safety analyzes the text for harmful content based on configured categories (Hate, Sexual, Violence, Self-Harm) and severity thresholds.
   • Custom blocklists can also be applied.
3. Decision Point:
   • If Safe: The message proceeds to the intended service (e.g., RAG, Direct Model Interaction, Image Generation).
   • If Unsafe: The message is blocked. The user receives a generic notification (or configured message). Details of the violation may be logged (if configured) and potentially viewable by users with the SafetyAdmin role.
4. Service Interaction (If Safe):
   • RAG / AI Search: The query is used to search Azure AI Search indexes (personal/group).
   • Direct Model Interaction: The message is sent directly to the Azure OpenAI GPT model.
   • Image Generation: The prompt is sent to the Azure OpenAI DALL-E model (if enabled).
   • Note: Responses from these services are typically not sent back through Content Safety by default in this flow, though Azure OpenAI itself has built-in content filtering.

Add your data (RAG Ingestion)

This workflow describes how documents uploaded via "Your Workspace" or "Group Workspaces" are processed for Retrieval-Augmented Generation.

1. User Uploads File(s):
   • User selects one or more supported files via the application UI (e.g., PDF, DOCX, TXT, MP4, MP3).
   • Files are sent to the backend application running on Azure App Service.
2. Initial Processing & Text Extraction:
   • The backend determines the file type.
   • The file is sent to the appropriate service for text extraction:
     • Azure AI Document Intelligence: For PDFs, Office Docs, Images (OCR). Extracts text, layout, tables.
     • Azure Video Indexer: For videos. Extracts audio transcript and frame OCR text (if enabled).
     • Azure Speech Service: For audio files. Extracts audio transcript (if enabled).
     • Internal Parsers: For plain text, HTML, Markdown, JSON, CSV.
3. Content Chunking:
   • The extracted text content is divided into smaller, manageable chunks based on file type and content structure.
   • Chunking strategies vary (see Advanced Chunking Logic under Latest Features) but aim for semantic coherence and appropriate size (~400-1200 words, depending on type), often with overlap between chunks to preserve context. Timestamps or page numbers are included where applicable.
4. Vectorization (Embedding):
   • Each text chunk is sent to the configured Embedding Model endpoint in Azure OpenAI.
   • The model generates a high-dimensional vector embedding (a numerical representation) for the semantic content of the chunk.
   • This process repeats for all chunks from the uploaded file(s).
5. Storage in Azure AI Search and Cosmos DB:
   • For each chunk, the following are stored in the appropriate Azure AI Search Index (simplechat-user-index or simplechat-group-index):
     • Chunk content (text).
     • Vector embedding.
     • Metadata: Parent document ID, user/group ID, filename, chunk sequence number, page number (if applicable), timestamp (if applicable), classification tags (if applicable), extracted keywords/summary (if applicable).
   • Metadata about the parent document (e.g., original filename, total chunks, upload date, user ID, group ID, document version, classification, processing status) is stored in Azure Cosmos DB.
   • Cosmos DB maintains the relationship between the parent document record and its constituent chunks stored in Azure AI Search.
6. Ready for Retrieval:
   • Once indexed, the document content is available for hybrid search (vector + keyword) when users toggle "Search Your Data" or perform targeted searches within workspaces.