Skip to content

RAG, Search, and Recommend

All three in-repo hosts share the same product scenarios. Embeddings and chat come from LM Studio (host concern); ZVec.NET stores and queries vectors.

RAG

  1. Seed / ingest documents (fixtures, EG FAQ CSV, optional FiQA)
  2. Retrieve by dense vector (+ citations, score-descending, near-dup dedupe)
  3. Ask with chat — Maui/AspNet prefer SSE streaming (POST /rag/ask/stream)

Suggested prompts: DemoPromptCatalog chips (EN + Arabic product + Egyptian CS FAQ). AspNet: GET /hints.

Semantic search over seeded corpora:

  • Fixtures
  • NFCorpus / Quora (download packs into gitignored samples/datasets/cache/ on startup when used)

Recommend

Similar-item style queries over:

  • T0 fixtures
  • MovieLens / Amazon Beauty (mass-embed; EmbeddingGemma EOS/SEP warnings in LM Studio are non-fatal)

Datasets

See samples/datasets/README.md. Seed via Maui buttons, Console commands, or AspNet POSTs (capped).

Restart-safe collections

Samples use factory.OpenOrCreate. Paths live under AppData (Maui) or temp (Console/AspNet). If an old folder looks empty or broken after an upgrade, delete the zvec-samples-* folders once and re-seed.

Smoke checklist

  • [ ] Console → status → models
  • [ ] rag seed-fixtures → ask (EN or Arabic chip)
  • [ ] rag seed-eg-faq → ask (Egyptian CS chip)
  • [ ] search seed-fixtures → search
  • [ ] recommend seed-fixtures → recommend
  • [ ] AspNet: GET /status, GET /hints, seed + query POSTs
  • [ ] Maui: restart twice — no CreateAndOpen crash; Status shows three doc counts

Full notes: samples/README.md.