Why

Trigram search misses "I like camping → outdoors group with hiking trips". Embeddings bridge that gap. Required for any quality conversational discovery ("find me a relaxed weekend hangout" → matches without exact-word overlap).

Scope (MVP)

• Add `pgvector` extension via AshPostgres
• Embedding column on Group (name + description) and Huddl (title + description)
• Background job to generate/refresh embeddings on create/update (AshOban)
• New Ash actions: `Group.semantic_search(query)` and `Huddl.semantic_search(query)` returning top-k by cosine similarity
• Choose embedding provider: Voyage (cheap), OpenAI text-embedding-3-small, or Anthropic when available
• Wire into `/groups` and `/huddlz` search as a hybrid (trigram + vector with rank fusion)

Open questions

• Hybrid ranking strategy (rank fusion vs. score blend)
• Embedding refresh policy on edits