docs: document self-hosted v0.0.5 model mixing bug and v0.0.7 resolution (#1450)

[Adityakk9031]

Resolves #1450.

Description

Documents the self-hosted v0.0.5 embedding model mixing bug in the embeddings documentation, explaining the root cause of why Japanese memories query and profile retrieval returned empty results, and notes that upgrading to v0.0.7 resolves the issue by enforcing a locked embedding plan.

Root Cause

In version v0.0.5, the embedding configuration was not locked. The server could mix different embedding models between document write paths (e.g. OpenAI text-embedding-3-small with 1536 dimensions) and memory query paths (e.g. local English default bge-base-en-v1.5), causing vector search cosine similarities to drop to near-zero.

• English Memories: Succeeded because full-text search (FTS) tokenized words based on spaces, allowing the hybrid search engine to fall back on lexical match.
• Japanese Memories: Failed completely because Japanese lacks space-separated words, causing SQLite FTS5 simple tokenizer to treat sentences as single tokens. The queries returned {"results":[],"total":0} because both the vector path (due to model mixing) and the lexical path (due to tokenization mismatch) failed.

Resolution

This was resolved in v0.0.7 by locking the embedding plan uniformly across all write and query paths.