⚡ Bolt: Cache regex context lexicon compilation in openmed

.jules/bolt.md

@@ -0,0 +1,3 @@
+## 2024-05-18 - Memoizing Regex Compilation in Deterministic Clinical NLP Layers
+**Learning:** In openmed, text processing pipelines are highly constrained to be deterministic and offline. However, regex compilation and pattern generation using `get_clinical_cue_lexicon` inside functions like `_compiled_context_lexicon` created severe performance bottlenecks when invoked repeatedly (e.g., millions of times over patient records), turning a sub-second task into a nearly minute-long CPU loop per bulk record. Memory reminds us that deterministic string evaluations must memoize regex compilation using `@functools.lru_cache`.
+**Action:** When implementing or fixing deterministic parsing or context engines inside `openmed.clinical` or similar pipeline components, apply `@functools.lru_cache(maxsize=None)` to methods building dictionaries or regex patterns from finite sets of lexical rules (e.g., lexicons by language). This makes downstream runtime evaluation near-instantaneous for subsequent hits.

openmed/openmed/clinical/context.py

@@ -35,6 +35,7 @@
 
 from __future__ import annotations
 
+import functools
 import re
 from collections.abc import Iterable, Iterator, Mapping, Sequence
 from dataclasses import dataclass, replace
@@ -154,6 +155,7 @@ class _CompiledContextLexicon:
     backward_context_cues: frozenset[str]
 
 
+@functools.lru_cache(maxsize=None)
 def _compiled_context_lexicon(language: str | None = None) -> _CompiledContextLexicon:
     lexicon = get_clinical_cue_lexicon(language)
     token_boundaries = lexicon.token_boundaries