Dom/t2a_v2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: mode hybride Ollama — gemma3:27b pour CPAM, 12b pour codage

Browse Source 
Le pipeline utilise désormais gemma3:12b (rapide) pour le codage CIM-10
et gemma3:27b (meilleur raisonnement) pour la contre-argumentation CPAM.
Configurable via OLLAMA_MODEL_CPAM et OLLAMA_TIMEOUT_CPAM.

Inclut aussi : traçabilité source/page DAS, niveaux CMA ATIH, sévérité,
page tracker PDF, améliorations fusion et filtres DAS.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
dom
2026-02-17 17:53:53 +01:00
parent 4ef42dd3d3
commit 01d47f3c4b
20 changed files with 1025 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
src/main.py
View File

@@ -14,7 +14,7 @@ from .config import ANONYMIZED_DIR, INPUT_DIR, OUTPUT_DIR, REPORTS_DIR, STRUCTUR
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.pdf_extractor import extract_text, extract_text_with_pages
from .extraction.trackare_parser import parse_trackare
from .medical.cim10_extractor import extract_medical_info
from .medical.ghm import estimate_ghm
@@ -38,8 +38,8 @@ def process_pdf(pdf_path: Path) -> list[tuple[str, DossierMedical, Anonymization
t0 = time.time()
logger.info("Traitement de %s", pdf_path.name)
# 1. Extraction texte
raw_text = extract_text(pdf_path)
# 1. Extraction texte avec pages
raw_text, page_tracker = extract_text_with_pages(pdf_path)
logger.info(" Texte extrait : %d caractères", len(raw_text))
# 2. Classification
@@ -82,7 +82,10 @@ def process_pdf(pdf_path: Path) -> list[tuple[str, DossierMedical, Anonymization
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 = extract_medical_info(
parsed, anonymized_text, edsnlp_result, use_rag=_use_rag,
page_tracker=page_tracker, raw_text=raw_text,
)
dossier.source_file = pdf_path.name
dossier.document_type = doc_type
logger.info(" DP%s : %s", part_label, dossier.diagnostic_principal)