Feature: Add semantic search via Qdrant + Embedding

[shukebeta]

Implementation Plan

Phase 1 – Infrastructure

• Set up Qdrant via Docker on local machine
• Set up Ollama with bge-m3 model (better CJK support)
• Add QdrantUrl and OllamaUrl to appsettings.json
• Add NuGet packages: Qdrant.Client, OllamaSharp

Phase 2 – Indexing

• Create IEmbeddingService wrapping Ollama
• Create IVectorSearchService wrapping Qdrant
• Chunking strategy: notes < 500 chars as single chunk; longer notes split by paragraph
• One-off backfill script for existing ~30k notes
• Hook into note create/update/delete to keep index in sync

Phase 3 – Search API

• Add semantic search path to existing search endpoint
• Return top N results by similarity score, filtered by userId
• Merge with Manticore results: keyword hits ranked first, semantic hits appended

Phase 4 – Resilience

• Graceful fallback to Manticore-only if Qdrant/Ollama is unreachable

Notes

• Qdrant payload must include userId — filter on every query, never expose
  other users' notes
• isPrivate notes should also be filtered appropriately
• Start with nomic-embed-text for quick validation; switch to bge-m3 if
  CJK quality needs improvement