Summary
Fresh PGLite brain initialized with ZeroEntropy zembed-1 at 2560 dimensions can become wedged when gbrain apply-migrations / migration schema replay runs.
The DB itself remains structurally intact, but the migration ledger records repeated partials and doctor reports a wedged migration. The failure appears to be the migrate/schema replay path using default 1536-dim gateway state instead of the configured embedding_dimensions: 2560, causing schema replay to emit the HNSW chunk index even though pgvector HNSW caps at 2000 dims.
Related but not identical: #1141. That issue is Postgres/Supabase schema v66 -> v67. This report is the PGLite fresh-install/apply-migrations surface.
Environment / config
Current upstream inspected while triaging:
master: 3a0e1116e76b1be45a136cf184d1a49d431b6b60VERSION: 0.36.1.0- Engine: PGLite
- Pages:
0
Config shape:
{
"engine": "pglite",
"database_path": "~/.gbrain/brain.pglite",
"embedding_model": "zeroentropyai:zembed-1",
"embedding_dimensions": 2560
}Observed
Doctor / state after failure:
schema_version remained current / unchangedconnection: Connectedpages: 0search_mode: balanced- config preserved
- migration ledger had repeated
partial entries for 0.11.0, later also 0.12.0
- doctor reported
WEDGED MIGRATION(s) after repeated partials
Representative underlying failure:
column cannot have more than 2000 dimensions for hnsw index
Expected
All schema replay / migration paths should apply the same HNSW policy as fresh schema generation:
- if embedding dimensions <= 2000: create
idx_chunks_embedding HNSW index
- if embedding dimensions > 2000: skip HNSW index and rely on exact vector scan
No migration should wedge a 2560-dim brain by trying to create the chunk HNSW index.
Suspected root
init --migrate-only loads config and calls engine.initSchema(), but does not appear to configure the AI gateway from config before schema generation.
PGLite initSchema() reads dimensions from the gateway and falls back to 1536 when the gateway is not configured. That makes getPGLiteSchema() apply the <=2000 HNSW branch even though the actual brain config/column is 2560.
This may also affect the in-process PGLite schema path in the open #1182 fix-wave unless that path configures the gateway before eng.initSchema().
Related
Workaround
Use zembed-1 at 1280 dimensions for PGLite fresh installs until the migrate/schema replay path is fixed.
Happy to close this as a duplicate of #1141 if maintainers prefer, but wanted to preserve the PGLite-specific repro surface.
Summary
Fresh PGLite brain initialized with ZeroEntropy
zembed-1at 2560 dimensions can become wedged whengbrain apply-migrations/ migration schema replay runs.The DB itself remains structurally intact, but the migration ledger records repeated partials and doctor reports a wedged migration. The failure appears to be the migrate/schema replay path using default 1536-dim gateway state instead of the configured
embedding_dimensions: 2560, causing schema replay to emit the HNSW chunk index even though pgvector HNSW caps at 2000 dims.Related but not identical: #1141. That issue is Postgres/Supabase schema v66 -> v67. This report is the PGLite fresh-install/apply-migrations surface.
Environment / config
Current upstream inspected while triaging:
master:3a0e1116e76b1be45a136cf184d1a49d431b6b60VERSION:0.36.1.00Config shape:
{ "engine": "pglite", "database_path": "~/.gbrain/brain.pglite", "embedding_model": "zeroentropyai:zembed-1", "embedding_dimensions": 2560 }Observed
Doctor / state after failure:
schema_versionremained current / unchangedconnection: Connectedpages: 0search_mode: balancedpartialentries for0.11.0, later also0.12.0WEDGED MIGRATION(s)after repeated partialsRepresentative underlying failure:
Expected
All schema replay / migration paths should apply the same HNSW policy as fresh schema generation:
idx_chunks_embeddingHNSW indexNo migration should wedge a 2560-dim brain by trying to create the chunk HNSW index.
Suspected root
init --migrate-onlyloads config and callsengine.initSchema(), but does not appear to configure the AI gateway from config before schema generation.PGLite
initSchema()reads dimensions from the gateway and falls back to 1536 when the gateway is not configured. That makesgetPGLiteSchema()apply the <=2000 HNSW branch even though the actual brain config/column is 2560.This may also affect the in-process PGLite schema path in the open #1182 fix-wave unless that path configures the gateway before
eng.initSchema().Related
Workaround
Use
zembed-1at 1280 dimensions for PGLite fresh installs until the migrate/schema replay path is fixed.Happy to close this as a duplicate of #1141 if maintainers prefer, but wanted to preserve the PGLite-specific repro surface.