Bug Description

_init() in kagent-feast-mcp/mcp-server/server.py performs unsynchronised
lazy initialisation. Under concurrent tool calls, multiple threads can observe
model is None simultaneously and each create a separate SentenceTransformer
instance, wasting memory and causing non-deterministic behaviour.

Root Cause

The check-then-act sequence has no lock, making it a classic TOCTOU race:

def _init():
    global model, client
    if model is None:          # thread A and B both see None here
        model = SentenceTransformer(EMBEDDING_MODEL)   # both initialise
    if client is None:
        client = MilvusClient(...)                     # both initialise

Impact

Redundant model loads under burst traffic increase memory pressure
Non-deterministic which instance ends up in the global, making behaviour hard to reason about in production

Proposed Fix

Apply the double-checked locking pattern with a module-level threading.Lock(): a fast unsynchronised check on the hot path, with a safe re-check under the lock to guarantee single initialisation.

I will submit a PR with this fix.