feat: Redis vector storage for PDF-based RAG

[Copilot]

Adds infrastructure for ingesting PDF documents into Redis vector indexes and surfacing retrieved context to agents via MCP tools. Agents like HeatingAgent and AppliancesAgent can now search ingested manuals (heatpump setup guides, Miele appliance PDFs) for relevant context when answering questions.

Vector store (raw Redis Search)

Microsoft.Extensions.VectorData.Redis doesn't exist on NuGet yet, so the implementation uses FT.CREATE/FT.SEARCH commands directly via StackExchange.Redis. HNSW indexing with configurable distance metric (default: cosine).

// Key schema
rag:{collection}:{chunkId}     // Hash — chunk content + embedding vector
rag:{collection}:idx           // FT.INDEX — vector similarity index
rag:meta:{collection}:{docId}  // Hash — document metadata

PDF ingestion pipeline

PdfDocumentIngestionService: PdfPig extraction → token-approximate chunking (configurable size/overlap) → batch embedding via IEmbeddingGenerator<string, Embedding<float>> → upsert to Redis.

Embedding generator factory

EmbeddingGeneratorFactory — mirrors the existing AgentExtensions.CreateAgent pattern for IChatClient. Supports Ollama (OllamaApiClient directly implements IEmbeddingGenerator), OpenAI, and AzureOpenAI. Local helper until CasCap.Common.AI adds a shared version.

MCP tools (RagMcpQueryService)

• SearchDocuments — KNN vector similarity search
• ListDocuments / GetDocumentInfo — collection metadata
• IngestDocument / RemoveDocument — on-demand management

Infrastructure

• docker-compose: redis → redis/redis-stack-server (enables Search module); ollama-init pulls nomic-embed-text (768d)
• RagIngestionBgService: auto-ingests PDFs from configured directory on startup
• Feature-gated: FeatureNames.Rag — opt-in, zero impact on existing features

Configuration

"RagConfig": {
  "EmbeddingProvider": "EdgeEmbedding",
  "IndexName": "rag-documents",
  "Dimension": 768,
  "ChunkSizeTokens": 512,
  "ChunkOverlapTokens": 50,
  "TopK": 5,
  "DocumentsPath": "/data/rag-documents",
  "AgentSources": {
    "HeatingAgent": [{ "CollectionName": "rag-documents", "TopK": 5 }],
    "AppliancesAgent": [{ "CollectionName": "rag-documents", "TopK": 3 }]
  }
}

Notes for follow-up

• EmbeddingGeneratorFactory should migrate to CasCap.Common.AI as a shared AgentExtensions.CreateEmbeddingGenerator
• RAG context middleware (IChatClient decorator that auto-injects retrieved chunks into agent prompts) is a natural next step
• When Microsoft.Extensions.VectorData.Redis ships on NuGet, RedisDocumentVectorStore can be simplified to use the official abstraction