⚡ Bolt: [memoize context lexicon regex compilation]

.jules/bolt.md

@@ -0,0 +1,3 @@
+## 2024-07-22 - [Missing memoization on regex compilation]
+**Learning:** Functions that generate compiled regex patterns on the fly (e.g., `_compiled_context_lexicon` in `openmed.clinical.context`) create severe performance bottlenecks when used repeatedly (e.g., during per-span text evaluations). Without `functools.lru_cache`, the regex compiler is continuously invoked, leading to a massive CPU overhead.
+**Action:** When working on text processing pipelines in openmed, always ensure deterministic regex compilation and lexicon generation steps are memoized (e.g., using `@functools.lru_cache`) to prevent unnecessary computation on repeated text segment evaluations.

openmed/openmed/clinical/context.py

@@ -36,6 +36,7 @@
 from __future__ import annotations
 
 import re
+import functools
 from collections.abc import Iterable, Iterator, Mapping, Sequence
 from dataclasses import dataclass, replace
 from datetime import date
@@ -154,6 +155,7 @@ class _CompiledContextLexicon:
     backward_context_cues: frozenset[str]
 
 
+@functools.lru_cache(maxsize=16)
 def _compiled_context_lexicon(language: str | None = None) -> _CompiledContextLexicon:
     lexicon = get_clinical_cue_lexicon(language)
     token_boundaries = lexicon.token_boundaries