How luismanez/collabdays-madrid-2025 works
Project type and purpose
• This is a microservice-style, full-stack RAG (Retrieval-Augmented Generation) demo:
– A .NET Core Web API (Abyx.Ai.Api) exposing chat completion endpoints with pluggable RAG memory.
– A background “ingestor” hosted service (Abyx.SharePoint.Lists.Ingestor) that pulls data from SharePoint via Microsoft Graph.
– A Copilot plugin integration layer (Abyx.Ai.CopilotPlugin) that registers the API with M365 Copilot.
File structure
• src/Abyx.Ai.Api
– Implements minimal API endpoints (GetChatCompletions).
– Uses a plugin architecture (Native and Semantic plugins) for custom RAG memory retrieval.
• src/Abyx.SharePoint.Lists.Ingestor
– A worker service that runs on a schedule to fetch SharePoint list items (Products) and load them into an in-memory or vector store.
• src/Abyx.Ai.CopilotPlugin
– Contains the ai-plugin.json, OpenAPI spec, adaptive cards, and manifest files that let M365 Copilot discover and call the API.
README for additional insights
• The solution demonstrates the RAG pattern with KernelMemory plus indexing SharePoint lists to supplement Copilot.
• Audience flow: SharePoint → Ingestor → Memory Store → API → Copilot.
Main components & their relationships
a. Data Source
– SharePoint Online Lists (Products list)
– Microsoft Graph API
b. Ingestion Layer
– Abyx.SharePoint.Lists.Ingestor (HostedService)
– Fetches list items, transforms to domain model (Product), pushes to RAG/Vector Store
c. Memory / Vector Store
– In-memory or persistent vector database (could be Azure Cognitive Search, Pinecone, or in-process KernelMemory)
d. API Layer
– Abyx.Ai.Api (Minimal Web API)
– Endpoint: /chat/completions uses RAG memory + LLM
– Plugin architecture:
• Native plugin (MyIpAddressPlugin)
• Semantic plugin (OurProductsPlugin) reading from the memory store
e. Integration Layer
– Abyx.Ai.CopilotPlugin app package
– openapi.yaml / ai-plugin.json / manifest.json
– Adaptive Cards + Declarative Agent config
f. External Consumer
– M365 Copilot (or any LLM client)
– Calls the registered plugin endpoint to augment chat with product info
Architectural patterns & design principles
• Microservices / Hosted services (separate API and ingestion)
• Plugin-based extensibility (IPlugin interface for RAG retrieval)
• Minimal Web API (lightweight endpoints)
• RAG pattern (combine vector search with LLM prompts)
• Declarative plugin registration (OpenAPI + ai-plugin.json for Copilot)
Essential elements
• Clear component names matching project folders (Abyx.SharePoint.Lists.Ingestor, Abyx.Ai.Api, Abyx.Ai.CopilotPlugin).
• Data flows annotated with request/response types (GetChatCompletionsRequest/Response, vector query).
• Plugin manifest (ai-plugin.json) pointing from Copilot to the Web API URL.