⚡ Bolt: Cache deterministic context lexicons for NLP processing

.jules/bolt.md

@@ -0,0 +1,3 @@
+## 2024-06-25 - Context Lexicon Regex Compilation Bottleneck
+**Learning:** Evaluated span assertions in the openmed NLP pipeline heavily depend on repeated deterministic cue compilations via `_compiled_context_lexicon`. This function constructs and formats numerous regexes for every evaluated axis (negation, temporality, certainty, etc). Doing this on every decision axis per text/span iteration caused massive redundant regex compilations and became a significant CPU bottleneck on the main thread.
+**Action:** Always memoize (e.g. `functools.lru_cache`) deterministic regex lexicons and configuration functions where inputs are purely hashable parameters like `language` strings.

openmed/openmed/clinical/context.py

@@ -39,6 +39,7 @@
 from collections.abc import Iterable, Iterator, Mapping, Sequence
 from dataclasses import dataclass, replace
 from datetime import date
+from functools import lru_cache
 from typing import Any, Literal
 
 from openmed.clinical.lexicons import (
@@ -154,6 +155,9 @@ class _CompiledContextLexicon:
     backward_context_cues: frozenset[str]
 
 
+# ⚡ Bolt: Cache deterministic context lexicons to avoid compiling the same
+# regex patterns thousands of times during span/text evaluations.
+@lru_cache(maxsize=32)
 def _compiled_context_lexicon(language: str | None = None) -> _CompiledContextLexicon:
     lexicon = get_clinical_cue_lexicon(language)
     token_boundaries = lexicon.token_boundaries