Release Notes
Added
- bge-base-en-v1.5 embedder (
--embedding-model bge-base): a 768-d Candle
BERT retriever (CLS pooling, query-only instruction prefix) selectable
alongside the default all-minilm. Assets are fetched on first use from a
pinned Hugging Face revision and verified by sha256, matching the existing
MiniLM download contract.
- Model-conditional default search variant: when
bge-base is selected and
--search-variant is not given, the retrieval default becomes dense-only
(hybrid fusion off). all-minilm and other models keep the hybrid-feature
default, and an explicit --search-variant always takes precedence.
Changed
- Warm workers report their embedding model and search variant in the ping
identity. A resident worker serving a different model/variant than the
client requests is now shut down and respawned (the same path as a
version/protocol skew), instead of silently answering with the wrong
embedder.
- Switching
--embedding-model on an existing store no longer wipes the
dense index. A persisted index built at a different vector dimension
(e.g. bge-base 768-d vs all-minilm 384-d) is incompatible with the active
model; memd now fails with a clear error naming both dimensions instead of
deleting the embedding cache and leaving an empty index. Set
MEMD_BACKFILL_HNSW_ON_STARTUP=1 to intentionally discard the dense index
and re-embed from segments under the new model. Same-dimension cache
corruption still deletes and rebuilds as before.
CandleEmbedder::with_config applies the model's required pooling strategy
(mean for MiniLM, CLS for bge-base), overriding any pooling carried on the
passed EmbeddingConfig — pooling is a property of the model's training
recipe, not a caller choice.
Install memd 1.4.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fmschulz/memd/releases/download/v1.4.0/memd-installer.sh | sh
Download memd 1.4.0
