Fix ai-service Dockerfile: bake in drug_entities.json, override its path

apps/ai-service/rag/routing.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import functools
 import re
 from dataclasses import dataclass, replace
 from difflib import SequenceMatcher
@@ -55,6 +56,20 @@ class CatalogDrugResolver:
         self._fuzzy_threshold = fuzzy_threshold
         self._ambiguity_margin = ambiguity_margin
 
+    # Measured live 2026-08-07: a single `resolve()` call over the real
+    # ~10,164-alias catalog costs ~0.65-0.7s, `suggest()` ~0.94-0.97s — both
+    # O(aliases) regex/SequenceMatcher work, pure functions of their
+    # arguments (only `self._aliases` et al, fixed at construction, feed
+    # them). `understanding.py`'s `_candidate_ids` calls both PER HISTORY
+    # LINE on every single turn — so the SAME already-seen history lines
+    # were being re-resolved from scratch every turn a conversation grew,
+    # ~1.6-1.7s of pure CPU per repeated line. A real user's ordinary
+    # multi-turn conversation was enough to exceed the 20s F-08 budget
+    # before the first Bedrock call ever ran, surfacing as a false
+    # "Dịch vụ đang gặp sự cố" — not a provider outage at all. Caching by
+    # exact input turns all but the newest turn's own text into a dict
+    # lookup on every subsequent call.
+    @functools.lru_cache(maxsize=4096)
     def resolve(self, query: str) -> DrugResolution:
         normalized_query = normalize_name(query)
         query_tokens = normalized_query.split()
@@ -140,6 +155,9 @@ class CatalogDrugResolver:
                 break
         return ordered
 
+    # See the comment on `resolve` above — same cost, same fix, same
+    # single-caller read-only usage (safe to hand back a cached list).
+    @functools.lru_cache(maxsize=4096)
     def suggest(
         self, query: str, k: int = 3, min_score: float = 0.5
     ) -> list[tuple[str, float]]: