⚡ Bolt: Memoize compiled context lexicon

.jules/bolt.md

@@ -0,0 +1,3 @@
+## 2024-07-25 - Context Lexicon Compilation Bottleneck
+**Learning:** Deterministic regex generation for clinical span lexicons (`_compiled_context_lexicon` in `openmed/clinical/context.py`) is expensive and is repeatedly invoked by various helpers (e.g. `is_negated`, `is_hypothetical`). Since the language argument is hashable, caching this generation avoids massive repetitive computation.
+**Action:** Always verify if computationally heavy dynamic regex generation based on statically loadable parameters (like a language code or fixed taxonomy) is being done per-span, and aggressively memoize these functions (`@functools.lru_cache`) to avoid re-compilation.

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 (
@@ -154,7 +155,13 @@ class _CompiledContextLexicon:
     backward_context_cues: frozenset[str]
 
 
+@lru_cache(maxsize=16)
 def _compiled_context_lexicon(language: str | None = None) -> _CompiledContextLexicon:
+    """Returns regex matchers derived from the context cues lexicon.
+
+    Memoized as cue sets can be large to iterate and regexes are expensive
+    to compile across many invocations.
+    """
     lexicon = get_clinical_cue_lexicon(language)
     token_boundaries = lexicon.token_boundaries
     return _CompiledContextLexicon(