Summary

Fresh PGLite installs can initialize the schema with the legacy OpenAI 1536-dim embedding column while the runtime gateway defaults to ZeroEntropy zembed-1 at 1280 dimensions. The first write/embed then fails with a dimension mismatch.

Observed behavior

On a fresh-ish local PGLite brain with no pages:

gbrain v0.37.1.0
config.json:
{
  "engine": "pglite",
  "database_path": "/home/opc/.gbrain/brain.pglite"
}

gbrain doctor --json reports the effective embedding provider as ZeroEntropy:

embedding_provider: ok / skipped
Model: zeroentropyai:zembed-1
connection: Connected, 0 pages

But a first gbrain put personal/goals.md fails when embedding:

expected 1536 dimensions, not 1280

The brain is empty, so this is not an old-data migration problem. The schema itself appears to have been initialized with vector(1536), while the active embedding path is producing 1280-dim ZeroEntropy vectors.

Why this looks like a default mismatch

The current runtime gateway defaults appear to be ZeroEntropy 1280:

// src/core/ai/gateway.ts
const DEFAULT_EMBEDDING_MODEL = 'zeroentropyai:zembed-1';
const DEFAULT_EMBEDDING_DIMENSIONS = 1280;

But several schema/init paths still appear to have legacy OpenAI defaults:

// src/core/pglite-schema.ts
export function getPGLiteSchema(dims: number = 1536, model: string = 'text-embedding-3-large'): string

Also found similar legacy fallbacks in PGLite/Postgres initialization paths:

src/core/pglite-engine.ts: let dims = 1536
src/core/postgres-engine.ts: let dims = 1536
src/core/schema-embedded.ts: embedding vector(1536)

This creates a split-brain default:

schema default: 1536 / OpenAI-style
runtime gateway default: 1280 / ZeroEntropy

Expected behavior

A new PGLite brain should be internally consistent.

Either:

1. The schema initializes to ZeroEntropy's default dimensions:

embedding_model = zeroentropyai:zembed-1
embedding_dimensions = 1280
content_chunks.embedding = vector(1280)

or:

2. If the schema intentionally remains 1536 for compatibility, the effective runtime embedding model/dimensions should also resolve to 1536 until the user explicitly runs a provider switch.

Suggested fix

Centralize the embedding defaults used by:

• gateway runtime config
• PGLite schema creation
• Postgres schema creation
• config seed rows
• doctor/schema dimension checks

At minimum, gbrain init --pglite should seed embedding_model and embedding_dimensions consistently with the schema it creates.

doctor should also catch this before first embed/write by comparing:

actual schema vector width
configured/effective embedding_dimensions
configured/effective embedding_model

Reproduction outline

gbrain init --pglite
# with no explicit --model / --embedding-model args
# configure ZEROENTROPY_API_KEY or otherwise let runtime resolve to zeroentropyai:zembed-1

gbrain put personal/goals.md < some-file.md

Actual:

expected 1536 dimensions, not 1280

Expected:

put succeeds, or doctor/init catches and fixes the mismatch before embedding

Impact

This breaks the first write for a new PGLite brain even when the brain has 0 pages, because the schema width and runtime embedding provider dimensions disagree.