Feature: AI-assisted transformation rule suggestions via column similarity

[peter7775]

Problem Statement

Transformation rules (config/*.yml) are written by hand today. On databases without declared foreign keys (common in legacy systems), it's hard to spot semantically equivalent columns across tables (e.g. customer_id ↔ client_no ↔ odberatel_id).

Proposed Solution

• Use column name/type embeddings (from the EmbeddingProvider port) and compute cosine similarity across all columns in the schema.
• For column pairs/groups above a similarity threshold, generate a candidate relationships section as YAML (dry-run output for the user to review and approve).
• Add a new CLI subcommand, e.g. sql-graph-visualizer analyze --suggest-rules, building on the existing analyze command.

Alternative Solutions

Pure structural heuristics (name/type matching without embeddings) — faster and has no embedding-provider dependency, but less accurate for differently named columns. Could serve as a fallback when no EmbeddingProvider is configured.

Use Case

Significantly shortens the most tedious part of onboarding a new database — hand-writing relationships rules — especially for databases without FK constraints.

Priority Level

Important

Implementation Ideas

• Start with an in-memory cosine similarity implementation in Go (no need to wait for the Neo4j/driver upgrade or native vector index) — trivial computationally at the scale of thousands of columns.
• Output a commented YAML diff/proposal rather than directly modifying the existing config.

Additional Context

Strongest candidate to implement first among the AI features — largely independent of the Neo4j upgrade issues since it can run fully in-memory.