pub const AI_DEFAULT_EMBEDDING_DIMS: usize = 768;Expand description
Default embedding vector dimensions used when the model path is unknown or empty (HTTP-only mode).
v0.10.0-dev reference value: 768 — set to match the production endpoint used during 0.10-dev development: Granite Embedding 278M multilingual on awa:8001 (768-dim vectors).
Operators using a different model MUST set this to match their endpoint:
- Granite Embedding 278M / nomic-embed-text / BGE-base → 768
- all-MiniLM-L6-v2 / BGE-small / GTE-small → 384
- stella_en_400M / jina-embeddings-v3 → 1024
Mismatch between this constant and the actual endpoint causes EmbeddingIndex (local usearch HNSW) to initialize with the wrong dimension, silently producing incorrect similarity scores. The HTTP embedding engine auto-detects dimensions from the endpoint probe and is unaffected, but the local HNSW index uses this constant as its fixed dimension at creation time.
Value 768 matches the Granite Embedding 278M production endpoint (awa:8001) used in v0.10.0-dev.
Operators using a different model (e.g. all-MiniLM-L6-v2 at 384-dim) must ensure their model
path is recognized by fxcp-ai/src/model_registry.rs so the correct dimension is returned.