perf: SentenceTransformer model reloaded on every search request

Bug Description

In both server/app.py and server-https/app.py, the SentenceTransformer
model is instantiated inside milvus_search(), meaning it is loaded from disk
on every single query:

def milvus_search(query: str, top_k: int = 5):
    encoder = SentenceTransformer(EMBEDDING_MODEL)  # ← re-loaded every call
    query_vec = encoder.encode(query).tolist()

Impact

Loading a transformer model takes 500ms–2s per call depending on hardware.
In an agentic RAG workflow where search_kubeflow_docs may be invoked
multiple times per conversation turn, this compounds into significant latency.

Proposed Fix

Move the encoder to a module-level singleton initialized once at startup:

# Module level — loaded once
encoder = SentenceTransformer(EMBEDDING_MODEL)

def milvus_search(query: str, top_k: int = 5):
    query_vec = encoder.encode(query).tolist()  # reuses loaded model
    ...

Files Affected

• server/app.py
• server-https/app.py

Start in server/app.py and server-https/app.py, reading milvus_search() and the current SentenceTransformer initialization. Make the encoder load once rather than per request, then verify that repeated searches in both server paths reuse it and still return query results without reloading the model.