Support configurable embedding dimensions for OpenAI-compatible providers

[sunyanliang2026]

Summary

• add OpenAI-compatible endpoint support in embedding client (OPENAI_BASE_URL / OPENAI_API_BASE and GBRAIN_OPENAI_* envs)
• make embedding model configurable via GBRAIN_EMBEDDING_MODEL
• make PGLite embedding vector dimension configurable via GBRAIN_EMBEDDING_DIMENSIONS
• validate returned embedding length against configured dimensions to fail fast on mismatch
• align chunk model defaults in engines with active embedding model

Why

Some OpenAI-compatible providers (for example siliconflow) expose embedding models that are not 1536 dimensions. Before this change, gbrain was pinned to vector(1536), which blocked semantic indexing on those providers even when /embeddings was otherwise working.

Implementation Notes

• PGLite schema SQL is now generated via pgliteSchemaSql() and reads model/dim defaults from env.
• Existing DBs keep their current schema; changing dimensions requires rebuilding index DB or explicit migration.
• Existing defaults remain unchanged (text-embedding-3-large, 1536) when env vars are not set.

Validation

• bun test test/pglite-engine.test.ts -t "searchVector returns empty when no embeddings"
• end-to-end local check against siliconflow-compatible setup:
  • sync/import succeeded
  • embed completed with Qwen/Qwen3-Embedding-0.6B + GBRAIN_EMBEDDING_DIMENSIONS=1024
  • gbrain stats showed full embedding coverage

[sunyanliang2026]

Closing at author request; keeping changes only in personal fork.