Bug: MCP server ignores embedding.model from .java-codebase-rag.yml

[HumanBean17]

Environment

• OS: macOS Darwin 25.5.0
• python-codebase-rag commit: 6b883d9 (master)

Problem

The MCP server (server.py) does not load .java-codebase-rag.yml YAML configuration for embedding.model. It always
falls back to the HuggingFace hub default sentence-transformers/all-MiniLM-L6-v2, causing search tool failures when
offline or behind a corporate proxy with self-signed certificates.

Reproduction

1. Create .java-codebase-rag.yml in the project root:
   embedding:
   model: /Users//Downloads/sentence-transformers:all-MiniLM-L6-v2
2. Configure MCP server in .mcp.json (no SBERT_MODEL env var):
   {
   "mcpServers": {
   "java-codebase-rag": {
   "command": ".venv/bin/java-codebase-rag-mcp",
   "env": {
   "LANCEDB_URI_PROJECT_ROOT": "/path/to/project"
   }
   }
   }
   }
3. Call the search tool — it fails with:
   RuntimeError: Cannot send a request, as the client has been closed.

Expected behavior

The MCP server should read embedding.model from .java-codebase-rag.yml and use the locally cached model, same as the
CLI does.

Root cause

The YAML config is resolved by resolve_operator_config() (config.py:248), which calls apply_to_os_environ() to set
SBERT_MODEL in the environment. However, this function is only called from CLI commands (init, reprocess, etc.), not
from the MCP server startup path.

The MCP server instead imports SBERT_MODEL from index_common.py:8 at module load time:

SBERT_MODEL = os.path.expandvars(
os.path.expanduser(os.environ.get("SBERT_MODEL", "sentence-transformers/all-MiniLM-L6-v2"))
)

Since SBERT_MODEL is not set in the MCP server's environment, it defaults to the HuggingFace hub name. When search_v2
runs (mcp_v2.py:903):

model_name = resolved_sbert_model_for_process_env(SBERT_MODEL)

It gets the hub name, and SentenceTransformer attempts to reach HuggingFace to check for adapter configs, which fails
behind corporate proxies or offline.

Affected tools

• search — BROKEN (requires SentenceTransformers / embeddings)
• find, resolve, describe, neighbors — unaffected (Kuzu graph only, no embedding model needed)

Workaround

Set SBERT_MODEL explicitly in .mcp.json env:
{
"mcpServers": {
"java-codebase-rag": {
"env": {
"SBERT_MODEL": "/Users//Downloads/sentence-transformers:all-MiniLM-L6-v2"
}
}
}
}

Suggested fix

In server.py, during startup, detect .java-codebase-rag.yml in the project root (using _project_root()), parse the
embedding.model key, and apply it to os.environ["SBERT_MODEL"] before any tool handler runs — mirroring what
resolve_operator_config().apply_to_os_environ() already does for the CLI path. The _pick_str /
_resolve_embedding_model_path functions in config.py already handle YAML precedence resolution and can be reused.

Key files

• index_common.py:7-10 — hardcoded default, import-time eval
• java_codebase_rag/config.py:248 — resolve_operator_config() (CLI-only path)
• java_codebase_rag/config.py:55-64 — resolved_sbert_model_for_process_env()
• server.py — MCP server startup (no YAML loading for embedding)
• mcp_v2.py:903 — where search_v2 reads the model name