get rid of local ocr, using gpt-4o instead

services/extractor.py

@@ -1,13 +1,15 @@
 import base64
 import binascii
+import io
+from PIL import Image
+from langchain_core.messages import HumanMessage, SystemMessage
 from typing import Callable, Type, Any, TypeVar
 from langchain_core.runnables import Runnable
 from langchain_core.prompts import ChatPromptTemplate
 from langchain_core.output_parsers import PydanticOutputParser
 from langchain_openai.chat_models import ChatOpenAI
 from pydantic import BaseModel
-
+from utils.parsers import process_profile, process_account, process_description ,process_passport
-from utils.parsers import process_profile, process_passport, process_account, process_description
 from validation.from_account import FromAccount
 from validation.from_passport import FromPassport
 from validation.from_profile import FromProfile
@@ -54,23 +56,43 @@ def extract_account(client_data: dict[str, Any])-> FromAccount:
 
 
 def extract_passport(client_data: dict[str, Any]) -> FromPassport:
-    passport_data = client_data.get("passport")
+    raw_file_data = client_data.get("passport")
 
-    prompt_template = (
+    if not raw_file_data:
-        "Extract the following information from the provided passport text.\n"
+        raise ValueError("Missing passport base64 data")
-        "Return only JSON matching this format:\n{format_instructions}\n\n"
+
-        "Pay special attention to the passport number\n"
+    try:
-        "Passport text:\n{processed_text}"
+        base64.b64decode(raw_file_data, validate=True)
+    except binascii.Error as e:
+        raise ValueError(f"Invalid base64 data: {e}")
+
+    # Décodage image
+    image_bytes = base64.b64decode(raw_file_data)
+    image = Image.open(io.BytesIO(image_bytes))
+
+    # Parser Pydantic
+    parser = PydanticOutputParser(pydantic_object=FromPassport)
+    format_instructions = parser.get_format_instructions()
+
+    # LLM gpt-4o
+    llm = ChatOpenAI(model="gpt-4o", temperature=0)
+
+    # Messages multimodaux
+    messages = [
+        SystemMessage(content="Tu es un assistant qui lit les passeports."),
+        HumanMessage(
+            content=[
+                {"type": "text", "text": f"Lis ce passeport et retourne les infos suivantes au format JSON :\n{format_instructions}"},
+                {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64," + raw_file_data}},
+            ]
         )
+    ]
 
-    result = __run_extraction_chain(
+    # Appel direct du LLM (hors prompt chain)
-        raw_file_data=passport_data,
+    result = llm.invoke(messages)
-        file_processor=process_passport,
-        pydantic_model=FromPassport,
-        prompt_template=prompt_template,
-    )
 
-    return result
+    # Parsing structuré
+    return parser.parse(result.content)
 
 
 def extract_profile(client_data: dict[str, Any]) -> FromProfile:

utils/parsers/passport_parser.py

@@ -1,42 +1,39 @@
 import base64
 import io
-from tempfile import NamedTemporaryFile
+from PIL import Image
-from PIL import Image, ImageEnhance
+from langchain_openai import ChatOpenAI
-import pytesseract
+from langchain.output_parsers import PydanticOutputParser
-from passporteye import read_mrz
+from langchain_core.prompts import ChatPromptTemplate
-import json
+from validation.from_passport import FromPassport
 
 def process_passport(passport_b64: str) -> str:
     """
     Traite le passport :
     - Décodage de l'image en base64.
-    - Application de l'OCR pour extraire le texte.
+    - Envoi à GPT-4o avec un prompt d'extraction structuré.
-
+    - Parsing structuré avec un modèle Pydantic.
-    :param passport_b64: Chaîne base64 représentant l'image du passport.
-    :return: Texte extrait de l'image.
     """
     image_bytes = base64.b64decode(passport_b64)
-    with NamedTemporaryFile(mode="wb") as tmp_img:
-        tmp_img.write(image_bytes)
-        with open(tmp_img.name, "rb") as read_img:
-            mrz_obj = read_mrz(read_img)
-
     image = Image.open(io.BytesIO(image_bytes))
-    enhancer = ImageEnhance.Contrast(image)
+
-    image = enhancer.enhance(2.0)  # 2.0 = double le contraste (1.0 = inchangé)
+    # Parser structuré basé sur le modèle FromPassport
-    tesseract_text = pytesseract.image_to_string(image, lang='eng')
+    parser = PydanticOutputParser(pydantic_object=FromPassport)
-    out_dict = {}
+
-    if not mrz_obj is None:
+    # Prompt + Instructions pour extraction
-        number_raw = str(mrz_obj.number)
+    prompt = ChatPromptTemplate.from_messages([
-        # It's not called a 'Hack'athon for nothing...
+        ("system", "Tu es un assistant d'extraction de données de passeport."),
-        number = number_raw.replace("B", "8")
+        ("human", "Voici l'image d'un passeport. Extrais les informations dans ce format :\n\n{format_instructions}"),
-        out_dict = {
+    ])
-            "country": mrz_obj.country,
+
-            "names": mrz_obj.names,
+    # LLM avec vision (GPT-4o)
-            "number": number,
+    llm = ChatOpenAI(model="gpt-4o", temperature=0)
-            "surname": mrz_obj.surname,
+
-            "mrz": mrz_obj.aux["text"],
+    chain = prompt | llm.with_structured_output(parser=parser)
-        }
+
-    out_dict["raw"] = tesseract_text
+    # Appel du LLM avec l'image en contexte
-    out = json.dumps(out_dict)
+    result = chain.invoke({
-    return out
+        "format_instructions": parser.get_format_instructions(),
+        "image": image,
+    })
+
+    return result.json()