Part of #390. Implements the decisions in #393.
What
- Message embedding: a detached step at
SESSION_END, independent of auto_reconcile, embeds the session's messages through MAGE's embeddings.text and stores the vectors in Memgraph. It costs no LLM, and the hook never waits on it.
- Reconciliation: embeds the entity names and edge source sentences (
r.text) it writes.
- Catch-up command: embeds anything missing, including failed steps.
- Model:
BAAI/bge-small-en-v1.5 by default ([recall] embedding_model). The model name and dimension are recorded with the vectors, so a change is detected rather than mixing vectors.
doctor: checks that the MAGE embeddings module exists and the model loads.- Docs and bootstrap: move from
memgraph/memgraph to memgraph-mage, with a 2 GiB minimum.
Done when
- A real session's messages, entities and edges have vectors.
- A model change is detected.
- On a plain Memgraph, embedding is reported unavailable instead of failing silently.
Part of #390. Implements the decisions in #393.
What
SESSION_END, independent ofauto_reconcile, embeds the session's messages through MAGE'sembeddings.textand stores the vectors in Memgraph. It costs no LLM, and the hook never waits on it.r.text) it writes.BAAI/bge-small-en-v1.5by default ([recall] embedding_model). The model name and dimension are recorded with the vectors, so a change is detected rather than mixing vectors.doctor: checks that the MAGEembeddingsmodule exists and the model loads.memgraph/memgraphtomemgraph-mage, with a 2 GiB minimum.Done when