Problem

The MCP server relies on ChromaDB's built-in default embedding function (all-MiniLM-L6-v2, 384 dimensions). There is no centralized embedding model configuration — each code path makes its own choice.

If a user ingests their palace with all-mpnet-base-v2 (768-dim) — which is what the ingest pipeline documentation and several internal paths assume — then every MCP query silently fails or returns garbage results because the query embeddings (384-dim) don't match the stored embeddings (768-dim).

This also affects anyone following issue #515 (GPU-accelerated embeddings via sentence-transformers) or #756 (OpenAI embeddings) — any non-default model used at ingest time will break MCP search.

Expected behavior

A single source of truth for which embedding model the palace was built with. The MCP server, CLI search, and all ingest paths should resolve to the same model.

Proposed fix

Centralized embedding config with 3-tier resolution:

1. Palace-level — {palace_path}/palace_meta.json stores the model used at build time
2. Global config — ~/.mempalace/config.json for user preference
3. Default — all-mpnet-base-v2 (768-dim)

mcp_server.py reads from this config instead of relying on ChromaDB's default. Ingest pipelines stamp the model into palace_meta.json at build time so the palace is self-describing.

This also resolves #860 (poor cross-room search with MiniLM) since mpnet produces significantly better embeddings for this use case — benchmarks show +3.5pp on LoCoMo R@10 with bge-large-en-v1.5 over MiniLM, and mpnet sits between the two.

Related issues

• Default embedding model (all-MiniLM-L6-v2) produces poor cross-room search results — large rooms dominate #860 — Default MiniLM produces poor cross-room search results
• feat: GPU-accelerated embeddings via optional sentence-transformers #515 — GPU-accelerated embeddings via sentence-transformers
• Add MemPalace OpenAI 3072-dim embedding support or something similar opensource. #756 — OpenAI 3072-dim embedding support
• Title: Search does not work for non-English languages (Russian, etc.) — default embedding model is English-only #712 — Default model is English-only
• Chunk size (800) exceeds embedding model token limit (256 tokens / ~512 chars) #390 — Chunk size exceeds embedding model token limit