System Architecture, RAG Pipeline & Multi-Model Integration
sajjad.ai is an enterprise-ready, multi-lingual AI software platform built from scratch. It empowers users to run conversational AI, interact with massive documentation sets via Retrieval-Augmented Generation (RAG), and execute smart content generation workflows in English, Arabic, and Urdu.
The system follows Clean Architecture principles, dividing concerns into Domain, Application, Infrastructure, and API presentation layers. This enables plug-and-play model updates and simple integrations with cloud databases or local servers.
Problem: Standard HTTP request-response patterns result in long load times and block pages when streaming tokens from OpenAI/DeepSeek.
Solution: Configured .NET Core Controller Web APIs with `Response.ContentType = "text/event-stream"` and serialized structured chunks on-the-fly. This pipes model responses instantly to the React frontend, allowing a smooth typing text animation with minimal latency.
Problem: Interacting with different vendor APIs (OpenAI, DeepSeek, and local Ollama) typically requires different clients, schemas, and payload specifications.
Solution: Architected a unified `ILLMService` interface using standard dependency injection. The system resolves execution targets at runtime based on the client configuration, abstracting chat histories, system prompt styling, and connection handshakes.
Problem: Standard vector databases search documents based on semantic similarity, but often lose the broader context of complex file sheets or PDF indexes.
Solution: Developed a custom C# parser that reads PDF, DOCX, and text datasets. Document streams are parsed with dynamic chunk overlapping (256-token window with 32-token overlap) and injected with file metadata tags (headers, page indices) prior to embedding generation. This significantly improves search precision in the vector search layer.
The platform is active and fully functional. You can review its live components directly at: