Refine scoring: preserve text structure for chunking and tune semantic calibration

services/scorer.py

@@ -4,7 +4,6 @@ from utilities.skills import (
     calculate_skill_overlap,
     extract_resume_skills,
     extract_required_skills_from_jd,
-    clean_text,
 )
 from services.feedback import generate_resume_feedback
 
@@ -60,12 +59,13 @@ def resume_score(resume_text: str, jd_text: str) -> dict:
     Orchestrates scoring → gap analysis → LLM feedback.
 
     Returns a dict matching ScoreResponse schema plus a 'summary' field.
-    """
-    resume_clean = clean_text(resume_text)
-    jd_clean     = clean_text(jd_text)
 
-    scores = final_ats_score(resume_clean, jd_clean)
-    gaps   = extract_gaps(resume_clean, jd_clean)
+    Note: text cleaning is intentionally left to each scoring/utility
+    function so that sentence boundaries (newlines, punctuation) are
+    preserved for semantic chunking before they are stripped.
+    """
+    scores = final_ats_score(resume_text, jd_text)
+    gaps   = extract_gaps(resume_text, jd_text)
 
     feedback = generate_resume_feedback(scores, gaps)
 

utilities/keyword_match.py

@@ -92,7 +92,7 @@ def calibrate_semantic_score(cosine: float) -> float:
     not 0.9+, so raw cosine understates good matches without calibration.
     """
     cosine = float(np.clip(cosine, 0.0, 1.0))
-    low, high = 0.32, 0.78
+    low, high = 0.20, 0.78
     scaled = (cosine - low) / (high - low) * 100.0
     return round(float(np.clip(scaled, 0.0, 100.0)), 2)