feat: découpage PDFs multi-dossiers (Trackare multi-épisodes, CRH concaténés)

Ajoute une étape de splitting entre extraction texte et parsing. Chaque chunk
est traité indépendamment par le pipeline existant, avec suffixe _partN en sortie.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/main.py

@@ -13,6 +13,7 @@ from .anonymization.anonymizer import Anonymizer
 from .config import ANONYMIZED_DIR, REPORTS_DIR, STRUCTURED_DIR, AnonymizationReport, DossierMedical
 from .extraction.document_classifier import classify
 from .extraction.crh_parser import parse_crh
+from .extraction.document_splitter import split_documents
 from .extraction.pdf_extractor import extract_text
 from .extraction.trackare_parser import parse_trackare
 from .medical.cim10_extractor import extract_medical_info
@@ -28,8 +29,11 @@ _use_edsnlp = True
 _use_rag = True
 
 
-def process_pdf(pdf_path: Path) -> tuple[str, DossierMedical, AnonymizationReport]:
-    """Traite un PDF : extraction → parsing → anonymisation → extraction CIM-10."""
+def process_pdf(pdf_path: Path) -> list[tuple[str, DossierMedical, AnonymizationReport]]:
+    """Traite un PDF : extraction → splitting → parsing → anonymisation → extraction CIM-10.
+
+    Retourne une liste de (texte_anonymisé, dossier, rapport) — un par dossier détecté.
+    """
     t0 = time.time()
     logger.info("Traitement de %s", pdf_path.name)
 
@@ -41,40 +45,53 @@ def process_pdf(pdf_path: Path) -> tuple[str, DossierMedical, AnonymizationRepor
     doc_type = classify(raw_text)
     logger.info("  Type de document : %s", doc_type)
 
-    # 3. Parsing
-    if doc_type == "trackare":
-        parsed = parse_trackare(raw_text)
-    else:
-        parsed = parse_crh(raw_text)
+    # 3. Splitting multi-dossiers
+    chunks = split_documents(raw_text, doc_type)
+    if len(chunks) > 1:
+        logger.info("  Découpage : %d dossiers détectés dans %s", len(chunks), pdf_path.name)
 
-    # 4. Anonymisation
-    anonymizer = Anonymizer(parsed_data=parsed)
-    anonymized_text = anonymizer.anonymize(raw_text)
-    report = anonymizer.report
-    report.source_file = pdf_path.name
-    logger.info(
-        "  Anonymisation : %d remplacements (regex=%d, ner=%d, sweep=%d)",
-        report.total_replacements,
-        report.regex_replacements,
-        report.ner_replacements,
-        report.sweep_replacements,
-    )
+    results: list[tuple[str, DossierMedical, AnonymizationReport]] = []
+    for i, chunk_text in enumerate(chunks):
+        part_label = f" [part {i+1}/{len(chunks)}]" if len(chunks) > 1 else ""
+        logger.info("  Traitement%s...", part_label)
 
-    # 5. Analyse edsnlp (optionnelle)
-    edsnlp_result = None
-    if _use_edsnlp:
-        edsnlp_result = _run_edsnlp(anonymized_text)
+        # 4. Parsing
+        if doc_type == "trackare":
+            parsed = parse_trackare(chunk_text)
+        else:
+            parsed = parse_crh(chunk_text)
 
-    # 6. Extraction médicale CIM-10
-    dossier = extract_medical_info(parsed, anonymized_text, edsnlp_result, use_rag=_use_rag)
-    dossier.source_file = pdf_path.name
-    dossier.document_type = doc_type
-    dossier.processing_time_s = round(time.time() - t0, 2)
-    logger.info("  DP : %s", dossier.diagnostic_principal)
-    logger.info("  DAS : %d, Actes : %d", len(dossier.diagnostics_associes), len(dossier.actes_ccam))
-    logger.info("  Temps de traitement : %.2fs", dossier.processing_time_s)
+        # 5. Anonymisation
+        anonymizer = Anonymizer(parsed_data=parsed)
+        anonymized_text = anonymizer.anonymize(chunk_text)
+        report = anonymizer.report
+        report.source_file = pdf_path.name
+        logger.info(
+            "  Anonymisation%s : %d remplacements (regex=%d, ner=%d, sweep=%d)",
+            part_label,
+            report.total_replacements,
+            report.regex_replacements,
+            report.ner_replacements,
+            report.sweep_replacements,
+        )
 
-    return anonymized_text, dossier, report
+        # 6. Analyse edsnlp (optionnelle)
+        edsnlp_result = None
+        if _use_edsnlp:
+            edsnlp_result = _run_edsnlp(anonymized_text)
+
+        # 7. Extraction médicale CIM-10
+        dossier = extract_medical_info(parsed, anonymized_text, edsnlp_result, use_rag=_use_rag)
+        dossier.source_file = pdf_path.name
+        dossier.document_type = doc_type
+        dossier.processing_time_s = round(time.time() - t0, 2)
+        logger.info("  DP%s : %s", part_label, dossier.diagnostic_principal)
+        logger.info("  DAS : %d, Actes : %d", len(dossier.diagnostics_associes), len(dossier.actes_ccam))
+
+        results.append((anonymized_text, dossier, report))
+
+    logger.info("  Temps total : %.2fs", time.time() - t0)
+    return results
 
 
 def _run_edsnlp(text: str):
@@ -252,10 +269,13 @@ def main(input_path: str | None = None) -> None:
         group_dossiers: list[DossierMedical] = []
         for pdf_path in pdfs:
             try:
-                anonymized_text, dossier, report = process_pdf(pdf_path)
+                pdf_results = process_pdf(pdf_path)
                 stem = pdf_path.stem.replace(" ", "_")
-                write_outputs(stem, anonymized_text, dossier, report, subdir=subdir)
-                group_dossiers.append(dossier)
+                multi = len(pdf_results) > 1
+                for part_idx, (anonymized_text, dossier, report) in enumerate(pdf_results):
+                    part_stem = f"{stem}_part{part_idx + 1}" if multi else stem
+                    write_outputs(part_stem, anonymized_text, dossier, report, subdir=subdir)
+                    group_dossiers.append(dossier)
             except Exception:
                 logger.exception("Erreur lors du traitement de %s", pdf_path.name)