schema-sync-agent-assisted

Schema Sync — Agent Assisted

Context

This system has four layers that break simultaneously when the database schema changes:

Layer	File / System	What breaks
YAML contract	agentic_system/schema/github_events.yaml	Contract no longer matches live DB
NL2SQL prompt	agentic_system/agents_core/nl2sql/prompts/system.md	Agent generates SQL with old predicates → 0 rows, no error
RAG prompt	agentic_system/agents_core/rag/prompts/system.md	Agent describes wrong field semantics to the user
ChromaDB chunks	schema_docs, qa_examples collections	RAG retrieves stale context with wrong field names

Repair procedure

Step 1 — Introspect live schema

Query ClickHouse to see what the database actually has:

SELECT name, type FROM system.columns WHERE table = 'github_events' ORDER BY position

Use connection settings from agentic_system/config.py (host, port, user, password, database).

Step 2 — Diff against YAML contract

Load agentic_system/schema/github_events.yaml and compare column names and types. Identify: removed columns, added columns, type changes.

Step 3 — Patch YAML contract

Update agentic_system/schema/github_events.yaml:

• Add entries for new columns with correct type and a clear post-migration description.
• Update the type field for any column whose type changed.
• Update descriptions for migration-sensitive columns (especially merge-related fields).

Step 4 — Patch agent prompts

In both system.md prompt files:

• Replace old field name references with new names.
• Replace old SQL predicates (e.g. field = 1) with new equivalents (e.g. field IS NOT NULL).
• Update any type annotations in the schema table within the prompt.
• Update state labels (pre-migration → post-migration).

Step 5 — Patch ChromaDB chunks

Connect to ChromaDB at http://localhost:8000 (settings from agentic_system/config.py). For each collection (schema_docs, qa_examples):

1. Get all items: POST /api/v2/tenants/default_tenant/databases/default_database/collections/{id}/get with {"include": ["documents", "metadatas"]}
2. Find items where metadata.stale == True.
3. Apply the same text replacements to the document content.
4. Re-embed using the deterministic hash function (64 dimensions, same algorithm as db/vectordb/init/seed_vectors.py).
5. Upsert back: POST .../upsert with updated documents, embeddings, and metadata (stale: false, schema_state: post_migration_synced).