Feature: Semantic schema search

[peter7775]

Problem Statement

Search/filtering in the graph visualization (Neovis.js) currently only matches on exact names. On older or inconsistently named databases (e.g. inv_hdr instead of invoices), users can't find relevant tables using natural language.

Proposed Solution

• During transformation, generate embeddings from table/column names + comments/data types via the EmbeddingProvider port and store them in the vector index.
• Add a GraphQL/REST endpoint for semantic queries (e.g. "customer payments") that returns the nearest nodes by cosine similarity.
• Wire this into the existing filter/search UI in the web layer (internal/interfaces/web/).

Alternative Solutions

Full-text search (Neo4j Lucene index) — simpler, but won't capture semantic similarity without exact word matches.

Use Case

A user working with a large legacy schema can quickly find relevant tables/columns without knowing exact names.

Priority Level

Nice to have

Implementation Ideas

• Extend schema/schema.graphqls with a new query type for semantic search.
• Generate embeddings in batch during the transform CLI command so normal runs aren't slowed down when AI is disabled.

Additional Context

Depends on: "Native vector index foundation", "EmbeddingProvider port".