0605-17:09

CSV/T11123/T11123.csv

@@ -0,0 +1,5 @@
+ÉMETTEUR,TYPE,DATE,CONTENU,ÉLÉMENTS VISUELS
+CLIENT,question,28/03/2025 15:00,Les parties d’ouvrage n’apparaissent plus.,"L'image montre une interface de gestion de chantier fonctionnelle, mais aucune mention explicite des ""parties d’ouvrage"" n'est visible."
+SUPPORT,réponse,28/03/2025 15:59,"Nous venons d'appliquer un correctif sur votre site, les parties de chantier sont de nouveau accessibles.",L'image confirme que l'interface est accessible avec des champs modifiables et des boutons actifs.
+CLIENT,question,28/03/2025 16:02,Je ne peux plus accéder à CBAO.
+SUPPORT,réponse,28/03/2025 16:06,Votre site est bien accessible à l'adresse suivante : https://nob.brg-lab.com/.,"L'image confirme que le client ""NORD OUEST BETON"" est bien visible dans la liste des laboratoires."

CSV/T11142/T11142.csv

@@ -0,0 +1,3 @@
+ÉMETTEUR,TYPE,DATE,CONTENU,ÉLÉMENTS VISUELS
+CLIENT,question,02/04/2025 15:52,Il est impossible de mettre des décimales sur la valeur de CO2/Tonne dans les caractéristiques des granulats de carrière.,Aucune image disponible
+SUPPORT,réponse,03/04/2025 07:44,Nous constatons bien ce dysfonctionnement. Un ticket a été ouvert auprès de notre équipe de développement. Vous serez automatiquement informé de sa résolution.,Aucune image disponible

CSV/T11143/T11143.csv

@@ -0,0 +1,4 @@
+ÉMETTEUR,TYPE,DATE,CONTENU,ÉLÉMENTS VISUELS
+CLIENT,question,03/04/2025 08:34,"Bonjour, Je ne parviens pas à accéder au l’essai au bleu. Merci par avance pour votre. Cordialement",Essai au bleu de méthylène de méthylène (MB) - NF EN 933-9 (02-2022)
+SUPPORT,réponse,03/04/2025 12:17,"Bonjour, Pouvez-vous vérifier si vous avez bien accès à la page suivante en l'ouvrant dans votre navigateur : https://zk1.brg-lab.com/ Voici ce que vous devriez voir affiché : Si ce n'est pas le cas, pouvez-vous me faire une capture d'écran de ce qui est affiché? Je reste à votre entière disposition pour toute information complémentaire. Cordialement,","Page d'accueil par défaut d'Apache Tomcat avec le message ""It works!"""
+CLIENT,information,03/04/2025 12:21,"Bonjour, Le problème s’est résolu seul par la suite. Je vous remercie pour votre retour. Bonne journée PS : l’adresse fonctionne",Essai au bleu de méthylène de méthylène (MB) - NF EN 933-9 (02-2022)

agents/RAG/agent_rag_responder.py

@@ -0,0 +1,39 @@
+# agents/RAG/agent_rag_responder.py
+
+from ..base_agent import BaseAgent
+
+class AgentRagResponder(BaseAgent):
+    def __init__(self, llm, ragflow):
+        """
+        Initialise l'agent avec un LLM et un client Ragflow.
+        """
+        super().__init__("AgentRagResponder", llm)
+        self.ragflow = ragflow
+
+    def executer(self, question: str, top_k: int = 5) -> str:
+        """
+        Effectue une recherche dans Ragflow et interroge le LLM avec le contexte.
+        """
+        # 1. Recherche les documents similaires
+        documents = self.ragflow.rechercher(question, top_k=top_k)
+
+        if not documents:
+            return "Aucun document pertinent trouvé dans la base de connaissance."
+
+        # 2. Construit le contexte à envoyer au LLM
+        contexte = "\n---\n".join([doc.get("content", "") for doc in documents])
+
+        # 3. Construit le prompt final
+        prompt = (
+            f"Voici des extraits de documents liés à la question suivante :\n\n"
+            f"{contexte}\n\n"
+            f"Réponds de manière précise et complète à la question suivante :\n{question}"
+        )
+
+        # 4. Interroge le LLM
+        reponse = self.llm.interroger(prompt)
+
+        # 5. Ajoute à l'historique (facultatif)
+        self.ajouter_historique(prompt, reponse)
+
+        return reponse

agents/llama_vision/agent_vision_ocr.bak2

@@ -0,0 +1,223 @@
+import os
+import json
+import logging
+from datetime import datetime
+from typing import Optional
+
+from ..base_agent import BaseAgent
+from ..utils.pipeline_logger import sauvegarder_donnees
+from utils.ocr_cleaner import clean_text_with_profiles  # AJOUT
+from utils.ocr_utils import extraire_texte_fr
+
+
+logger = logging.getLogger("AgentVisionOCR")
+
+class AgentVisionOCR(BaseAgent):
+    """
+    Agent LlamaVision qui extrait du texte (OCR avancé) depuis une image.
+    """
+
+    def __init__(self, llm):
+        super().__init__("AgentVisionOCR", llm)
+
+        # Configuration des paramètres du modèle
+        self.params = {
+            "stream": False,
+            "seed": 0,
+            #"stop_sequence": [],
+            "temperature": 1.3,
+            #"reasoning_effort": 0.5,
+            #"logit_bias": {},
+            "mirostat": 0,
+            "mirostat_eta": 0.1,
+            "mirostat_tau": 5.0,
+            "top_k": 35,
+            "top_p": 0.85,
+            "min_p": 0.06,
+            "frequency_penalty": 0.15,
+            "presence_penalty": 0.1,
+            "repeat_penalty": 1.15,
+            "repeat_last_n": 128,
+            "tfs_z": 1.0,
+            "num_keep": 0,
+            "num_predict": 2048,
+            "num_ctx": 16384,
+            #"repeat_penalty": 1.1,
+            "num_batch": 2048,
+            #"mmap": True,
+            #"mlock": False,
+            #"num_thread": 4,
+            #"num_gpu": 1                   
+        }
+
+        # Prompt OCR optimisé
+        self.system_prompt = ("""You are tasked with performing an exhaustive OCR extraction on a technical or administrative web interface screenshot.
+
+GOAL: Extract **every legible piece of text**, even partially visible, faded, or cropped. Structure your output for clarity. Do not guess, but always report what is visible.
+
+FORMAT USING THESE CATEGORIES:
+
+1. PAGE STRUCTURE
+- Page titles
+- Interface headers or section labels
+- Navigation bars or visible URLs
+
+2. IDENTIFIERS & DATA
+- Operator or user names
+- Sample IDs, test references
+- Materials, dates, batch numbers
+
+3. INTERFACE ELEMENTS (MANDATORY SCAN)
+- Button labels (e.g., RAZ, SAVE)
+- Tabs (e.g., MATERIAL, OBSERVATIONS)
+- Sidebars, form field labels
+
+4. SYSTEM MESSAGES
+- Connection or server errors
+- Domains, IP addresses, server notices
+
+5. METADATA
+- Standard references (e.g., "NF EN ####-#")
+- Version numbers, document codes, timestamps
+
+6. UNCLEAR / CROPPED TEXT
+- Logos, partial lines (use “[...]” for truncated)
+- Background/faded elements, labels not fully legible
+
+RULES:
+- Preserve punctuation, case, accents exactly.
+- Include duplicates if text appears more than once.
+- Never skip faint or partial text; use “[...]” if incomplete.
+- Even if cropped, report as much as possible from any UI region.
+
+This prompt is designed to generalize across all web portals, technical forms, or reports. Prioritize completeness over certainty. Do not ignore UI components or system messages.
+""")
+
+        self._configurer_llm()
+        self.resultats = []
+        self.images_traitees = set()
+        logger.info("AgentVisionOCR initialisé avec prompt amélioré.")
+
+    def _configurer_llm(self):
+        if hasattr(self.llm, "prompt_system"):
+            self.llm.prompt_system = self.system_prompt
+        if hasattr(self.llm, "configurer"):
+            self.llm.configurer(**self.params)
+
+    def _extraire_ticket_id(self, image_path):
+        if not image_path:
+            return "UNKNOWN"
+        segments = image_path.replace('\\', '/').split('/')
+        for segment in segments:
+            if segment.startswith('T') and segment[1:].isdigit():
+                return segment
+            if segment.startswith('ticket_T') and segment[8:].isdigit():
+                return 'T' + segment[8:]
+        return "UNKNOWN"
+
+    def executer(self, image_path: str, ocr_baseline: str = "", ticket_id: Optional[str] = None) -> dict:
+        image_path_abs = os.path.abspath(image_path)
+        image_name = os.path.basename(image_path)
+
+        if image_path_abs in self.images_traitees:
+            logger.warning(f"[OCR-LLM] Image déjà traitée, ignorée: {image_name}")
+            print(f" AgentVisionOCR: Image déjà traitée, ignorée: {image_name}")
+            return {
+                "extracted_text": "DUPLICATE - Already processed",
+                "image_name": image_name,
+                "image_path": image_path_abs,
+                "ticket_id": ticket_id or self._extraire_ticket_id(image_path),
+                "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
+                "source_agent": self.nom,
+                "is_duplicate": True
+            }
+
+        self.images_traitees.add(image_path_abs)
+        logger.info(f"[OCR-LLM] Extraction OCR sur {image_name}")
+        print(f" AgentVisionOCR: Extraction OCR sur {image_name}")
+
+        ticket_id = ticket_id or self._extraire_ticket_id(image_path)
+
+        try:
+            if not os.path.exists(image_path):
+                raise FileNotFoundError(f"Image introuvable: {image_path}")
+
+            if not hasattr(self.llm, "interroger_avec_image"):
+                raise RuntimeError("Le modèle ne supporte pas l'analyse d'images.")
+            # Etape 1: OCR br
+            # Interroger le modèle
+            response = self.llm.interroger_avec_image(image_path, self.system_prompt)
+
+            if not response or "i cannot" in response.lower():
+                raise ValueError("Réponse vide ou invalide du modèle")
+
+            cleaned_text = clean_text_with_profiles(response.strip(), active_profiles=("ocr",))
+
+            model_name = getattr(self.llm, "pipeline_normalized_name", 
+                        getattr(self.llm, "modele", "llama3-vision-90b-instruct"))
+            model_name = model_name.replace(".", "-").replace(":", "-").replace("_", "-")
+
+            result = {
+                "extracted_text": cleaned_text,
+                "image_name": image_name,
+                "image_path": image_path_abs,
+                "ticket_id": ticket_id,
+                "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
+                "source_agent": self.nom,
+                "model_info": {
+                    "model": model_name,
+                    **self.params
+                }
+            }
+
+            self.resultats.append(result)
+            logger.info(f"[OCR-LLM] OCR réussi ({len(cleaned_text)} caractères) pour {image_name}")
+            return result
+
+        except Exception as e:
+            error_result = {
+                "extracted_text": "",
+                "image_name": image_name,
+                "image_path": image_path_abs,
+                "ticket_id": ticket_id or "UNKNOWN",
+                "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
+                "source_agent": self.nom,
+                "error": str(e),
+                "model_info": {
+                    "model": getattr(self.llm, "pipeline_normalized_name", "llama3-vision-90b-instruct"),
+                    **self.params
+                }
+            }
+            self.resultats.append(error_result)
+            logger.error(f"[OCR-LLM] Erreur OCR pour {image_name}: {e}")
+            return error_result
+
+    def sauvegarder_resultats(self, ticket_id: str = "T11143") -> None:
+        if not self.resultats:
+            logger.warning("[OCR-LLM] Aucun résultat à sauvegarder")
+            return
+
+        resultats_dedupliques = {}
+        for resultat in self.resultats:
+            image_path = resultat.get("image_path")
+            if not image_path:
+                continue
+            if image_path not in resultats_dedupliques or \
+               resultat.get("timestamp", "") > resultats_dedupliques[image_path].get("timestamp", ""):
+                resultats_dedupliques[image_path] = resultat
+
+        resultats_finaux = list(resultats_dedupliques.values())
+
+        try:
+            logger.info(f"[OCR-LLM] Sauvegarde de {len(resultats_finaux)} résultats")
+            sauvegarder_donnees(
+                ticket_id=ticket_id,
+                step_name="ocr_llm",
+                data=resultats_finaux,
+                base_dir=None,
+                is_resultat=True
+            )
+            self.resultats = []
+            self.images_traitees = set()
+        except Exception as e:
+            logger.error(f"[OCR-LLM] Erreur sauvegarde résultats: {e}")

prompts/test_prompt_ocr4.txt

@@ -0,0 +1,231 @@
+Prompt de base:
+You are tasked with performing a high-precision OCR extraction on a partially cropped screenshot of a technical web interface.
+
+GOAL: Extract **all visible and partially visible text** from the image, no matter how small, faint, or cropped. Structure the output clearly to maximize usability.
+
+FORMAT THE OUTPUT AS FOLLOWS:
+
+---
+
+1. PAGE STRUCTURE
+* Page title or headers (e.g., test names, document references)
+* Section labels or grouping titles
+* URL or navigation path (if visible)
+
+2. LABORATORY DATA
+* Sample references or IDs (e.g., 25-00075)
+* Material descriptions (e.g., Sable 0/2 C)
+* Dates, operators (e.g., 02/04/2025 – BOLLÉE Victor)
+
+3. UI ELEMENTS
+* Sidebar text, tabs, and form labels
+* Buttons (e.g., RAZ, IMPRIMER, ENREGISTRER)
+* Navigation or functional labels (e.g., ESSAI, SMQ)
+
+4. SYSTEM MESSAGES & ERRORS
+* Any status or error message (e.g., "Impossible de trouver l'adresse IP du serveur de zkt1.brg-lab.com")
+* Domain names, IPs, server paths
+
+5. METADATA
+* Document standards (e.g., NF EN 933-9)
+* Version/date (e.g., 02-2022)
+* Any hidden identifiers, codes, or repeated numbers
+
+6. PERIPHERAL / UNCERTAIN TEXT
+* Logos (e.g., BRG-LAB), watermarks, faded/blurred text
+* Partially cropped lines: use “[...]” to indicate incomplete text
+
+RULES:
+- Do NOT interpret or translate.
+- Preserve all punctuation, accents, and casing.
+- Include duplicated text if present multiple times.
+- Mark unclear/partial lines explicitly.
+- Report even faint or background text.
+
+This is a high-precision extraction. Output clean bullet points grouped by section. Stay fully factual.
+
+Résultats:
+🔵 Résultat pour l'image output/ticket_T11143/T11143_20250422_084617/attachments/image_145435.png:
+Texte extrait:
+### 1\. Page Structure
+
+\- **Page title**: Essai au bleu de méthylène \(MB\) - NF EN 933-9 \((02\)-2022)
+
+\- **Header/Section labels**:
+
+\* BRG-LAB logo (top left)
+
+\* "Fournisseur Labo:" label with dropdown menu (right side)
+
+\- URL/path: zkt1.brg-lab.com (from the error message at the bottom) [](https://zk1.brg-lab.com)
+
+
+
+### 2\. Laboratory Data
+
+\- Sample reference: 25-00075 
+
+\- Material description: Sable 0/2 C
+
+\- Date and operator: 02/04/2025 - BOLLÉE Victor 
+
+
+
+### 3\. UI Elements
+
+\- Sidebar text/tabs/form labels:
+
+
+
+\+ **Tab bar**:
+
+\# Essai | Matériel | Observations | SMQ
+
+
+
+\* Form fields:
+
+    \+ "ESSAI"
+
+        - N° de fiche \(=\) ... (partially visible)
+
+    \+ "MATERIEL" / "Matériau": Sable 0/2 C
+
+    \+ "OBSERVATIONS" with empty field
+
+    \+ "SMQ" with an empty dropdown menu
+
+
+
+\* Buttons: RAZ, IMPRIMER, ENREGISTRER 
+
+
+
+### 4\. System Messages & Errors 
+
+\- Status/error message at the bottom of the page:
+
+Impossible de trouver l'adresse IP du serveur zk1.brg-lab.com.
+
+
+
+### 5\. Metadata  
+
+\- Document standard/version/date: NF EN 933-9 \(02\)-2022 
+
+
+
+### 6. Peripheral/Uncertain Text 
+
+
+
+\- BRG-LAB logo (top left)
+
+\- Faded/blurred text near buttons (right side) - partially readable words include "[...]tél", "dossier", and parts of what seems to be a form submission button or similar ("[...]" )
+
+
+Prompt corrigé sans éléments spécifiques:
+"""You are tasked with performing a high-precision OCR extraction on a partially cropped screenshot of a technical web interface.
+
+GOAL: Extract **all visible and partially visible text** from the image, no matter how small, faint, or cropped. Structure the output clearly to maximize usability.
+
+FORMAT THE OUTPUT AS FOLLOWS:
+
+---
+
+1. PAGE STRUCTURE
+* Page title or headers (e.g., test names, document references)
+* Section labels or grouping titles
+* URL or navigation path (if visible)
+
+2. LABORATORY DATA
+* Sample references or IDs 
+* Material descriptions 
+* Dates, operators 
+
+3. UI ELEMENTS
+* Sidebar text, tabs, and form labels
+* Buttons 
+* Navigation or functional labels 
+
+4. SYSTEM MESSAGES & ERRORS
+* Any status or error message 
+* Domain names, IPs, server paths
+
+5. METADATA
+* Document standards 
+* Version/date )
+* Any hidden identifiers, codes, or repeated numbers
+
+6. PERIPHERAL / UNCERTAIN TEXT
+* Logos (e.g., BRG-LAB), watermarks, faded/blurred text
+* Partially cropped lines: use “[...]” to indicate incomplete text
+
+RULES:
+- Do NOT interpret or translate.
+- Preserve all punctuation, accents, and casing.
+- Include duplicated text if present multiple times.
+- Mark unclear/partial lines explicitly.
+- Report even faint or background text.
+
+This is a high-precision extraction. Output clean bullet points grouped by section. Stay fully factual.
+"""
+Résultats:
+🔵 Résultat pour l'image output/ticket_T11143/T11143_20250422_084617/attachments/image_145435.png:
+Texte extrait:
+### **PAGE STRUCTURE**
+
+#### *Page title or headers*
+
+*   Essai au bleu de méthylène (MB) - NF EN 933-9 (02-2022)
+*   BOLLEE Victor
+*   PRELEVEUR: 25/06/75
+*   LABORATOIRE DE CONTROLE ET D'ANALYSE DES MATERIAUX
+
+#### *Section labels or grouping titles*: None visible.
+
+
+
+### **LABORATORY DATA**
+
+*   Sample references or IDs: 
+    *   N/A, no samples were referenced in the image provided.
+*   Material descriptions:
+    *   Echantillon n° 25-00073 réceptionné le 04/05/2025 par BOLLÉE Victo
+        (sample number partially obscured due to cropping.)
+    *   Matériau Sable C7C - CARRETERE ADCEG
+*   Dates and operators:
+
+    Date not legible; operator unclear.
+
+    *   Prélevée la 02/06/2025 p
+
+### **UI ELEMENTS**
+
+
+
+#### \_Sidebar text, tabs, and form labels:
+
+*   ESSAI
+*   MATÉRIEL
+*   PORTFOLIO
+
+#### Buttons: None visible.
+
+
+
+### **SYSTEM MESSAGES & ERRORS**
+
+None apparent from the image provided. The partial nature of the screenshot precludes any conclusive assessment in this regard.
+
+
+
+### PERIPHERAL / UNCERTAIN TEXT:
+
+
+
+Logo or watermark: BRG-LAB (cropped due to formatting)
+
+
+
+Note: This extraction strictly adheres to visible content without interpretation. Some fields are left blank where data is partially obscured by cropping or not discernible from the given snapshot.

prompts/test_prompt_ocr5.txt

@@ -0,0 +1,306 @@
+You are tasked with performing a high-precision OCR extraction on a screenshot of a technical or administrative web interface.
+
+GOAL: Extract all visible and partially visible text — no matter how small, faint, or cropped. Remain strictly factual. Do not interpret, guess, or reword.
+
+📄 FORMAT THE OUTPUT USING THESE CATEGORIES:
+
+---
+
+1. PAGE STRUCTURE
+* Page title(s)
+* Section or interface headers
+* Visible URLs, tabs, or menu paths
+
+2. DATA & IDENTIFIERS
+* Sample codes, test references, user/operator names
+* Material or item descriptions
+* Dates, times, unique identifiers
+
+3. INTERFACE ELEMENTS
+* Button labels
+* Tab names
+* Sidebar/menu content
+* Field or dropdown labels
+
+4. SYSTEM MESSAGES & ERRORS
+* Status messages, warnings, or connection errors
+* Domain names, IPs, server notices
+
+5. METADATA
+* Version numbers, standard references, document codes
+* Any duplicated text or footer content
+
+6. UNCLEAR OR CROPPED TEXT
+* Logos, watermarks, truncated words or symbols
+* Use “[...]” to mark incomplete or partially cropped text
+
+---
+
+RULES:
+- Do not translate or paraphrase.
+- Preserve original casing, spelling, punctuation.
+- Include repeated elements as they appear.
+- Report faint or background text if legible.
+- Leave blank sections if no relevant text is found.
+
+This prompt is designed to work across a wide range of web interfaces, dashboards, and structured forms. Output clearly grouped bullet points per section.
+
+Résultats:
+🔵 Paramètres actifs LLM:
+{
+  "temperature": 1.3,
+  "top_p": 0.85,
+  "presence_penalty": 0.1,
+  "frequency_penalty": 0.15,
+  "stop": [],
+  "stream": false,
+  "n": 1,
+  "seed": 0,
+  "mirostat": 0,
+  "mirostat_eta": 0.1,
+  "mirostat_tau": 5.0,
+  "top_k": 35,
+  "min_p": 0.06,
+  "repeat_penalty": 1.15,
+  "repeat_last_n": 128,
+  "tfs_z": 1.0,
+  "num_keep": 0,
+  "num_predict": 2048,
+  "num_ctx": 16384,
+  "num_batch": 2048
+}
+ AgentVisionOCR: Extraction OCR sur image_145435.png
+
+🔵 Résultat pour l'image output/ticket_T11143/T11143_20250422_084617/attachments/image_145435.png:
+Texte extrait:
+### 1) PAGE STRUCTURE 
+
+*   **Page title:** Essai au bleu de méthylène (MB)
+*   **Section/Interface headers:**
+    *   BRG-LAB
+*   **Visible URLs, tabs, menu paths:**
+    *   No visible URLs are displayed in the image.
+
+### 2) DATA & IDENTIFIERS 
+
+*   **Sample codes/test references/user/operator names:**
+    *   BOLLEE Victor 
+    *   Echantillon n° 25-00075 réceptionné le 02/04/2025 par BOLLEE Victor.
+*   **Material/item descriptions:**
+    *   Blue Methylene MB Testing Report
+
+**Note:** There is no mention of dates or times within the given page snippet. Additionally, there seems to be no unique identifiers on this web page interface screenshot provided.
+
+### 3) INTERFACE ELEMENTS  
+
+This section cannot be determined based on the information provided in the image.
+
+### 4) SYSTEM MESSAGES & ERRORS 
+
+There are no system messages or errors present in this snapshot as it only provides a view of what appears to be an itemized report rather than any kind of error message.
+
+### 5) METADATA  
+
+No metadata elements can be identified from the image alone such as version numbers standard references document codes etcetera without having access full context surrounding these items listed under "BRG-LAB". As far we know all text shown here pertains directly related subject matter topic discussed above thus falling outside scope defined categories listed above according guidelines outlined prompt.
+
+Nouveau prompt:
+You are tasked with performing a high-precision OCR extraction on a screenshot of a technical or administrative web interface.
+
+GOAL: Extract all visible and partially visible text — no matter how small, faint, or cropped. Remain strictly factual. Do not interpret, guess, or reword.
+
+📄 FORMAT THE OUTPUT USING THESE CATEGORIES:
+
+---
+
+1. PAGE STRUCTURE
+* Page titles (e.g., test names or document references)
+* Interface headers, menus or section titles
+* Navigation breadcrumbs, web paths, or visible URLs
+
+2. DATA & IDENTIFIERS
+* Sample numbers, user names, dates
+* Material types or item descriptions
+* Unique codes or test references
+
+3. INTERFACE ELEMENTS
+* Button labels (e.g., RESET, PRINT, SAVE)
+* Tabs or navigation elements (e.g., RESULTS, MATERIAL, OBSERVATIONS)
+* Form field labels or sidebar sections
+
+4. SYSTEM MESSAGES & ERRORS
+* Status or connection errors (e.g., “unable to connect”, IP/domain errors)
+* Visible domain names, system paths
+
+5. METADATA
+* Document codes, version numbers, dates (e.g., standard formats like “NF EN ####-#”)
+* Any repeated or footer elements
+
+6. UNCLEAR / CROPPED TEXT
+* Logos, watermarks, or faded UI elements
+* Truncated words or text: use “[...]” to indicate partial content
+
+---
+
+RULES:
+- Do not interpret or translate.
+- Preserve all accents, casing, punctuation.
+- Include all duplicates or repeated labels.
+- Capture system messages even if incomplete or faded.
+- Leave a section empty only if no corresponding text is found.
+
+Output a clean, bullet-pointed list per section. Be exhaustive and literal.
+
+Résultats:
+🔵 Paramètres actifs LLM:
+{
+  "temperature": 1.3,
+  "top_p": 0.85,
+  "presence_penalty": 0.1,
+  "frequency_penalty": 0.15,
+  "stop": [],
+  "stream": false,
+  "n": 1,
+  "seed": 0,
+  "mirostat": 0,
+  "mirostat_eta": 0.1,
+  "mirostat_tau": 5.0,
+  "top_k": 35,
+  "min_p": 0.06,
+  "repeat_penalty": 1.15,
+  "repeat_last_n": 128,
+  "tfs_z": 1.0,
+  "num_keep": 0,
+  "num_predict": 2048,
+  "num_ctx": 16384,
+  "num_batch": 2048
+}
+ AgentVisionOCR: Extraction OCR sur image_145435.png
+
+🔵 Résultat pour l'image output/ticket_T11143/T11143_20250422_084617/attachments/image_145435.png:
+Texte extrait:
+### Page Structure
+Page title: Essai au bleu de méthylène (MB) - NF EN 933-9 (02-2022)
+Navigation breadcrumbs/path: none visible
+
+### Data & Identifiers 
+Sample numbers/usernames/dates/material types/item descriptions/unique codes/test references:
+none are clearly readable in the provided image snippet; however, it appears there might be data within tables or sections not fully captured in this view.
+
+### Interface Elements
+Button labels/tabs/navigation elements/form field labels/sidebar sections: None explicitly identifiable from the given portion of the interface screenshot.
+
+### System Messages and Errors 
+Status/connection errors/domain names/system paths: None evident from the visible part of the page.
+
+### Metadata  
+Document codes/version numbers/dates/repeated/footer elements:
+NF EN 933-9 is mentioned which suggests a European standard reference but lacks specifics like date without further context. No version number, repeated text, or footer content is visible due to cropping.
+
+### Unclear/Cropped Text   
+Logos/watermarks/faded UI elements/truncated words/text with “[...]” indicating partial content:
+Due to significant portions being cropped out by the edges of this image snippet, several pieces of information are truncated including what seems to be a document title/standard reference and possibly interface labels or data entries. The full text for these items cannot be determined from the provided visual alone.
+
+Autre prompt:
+You are tasked with performing an exhaustive OCR extraction on a technical or administrative web interface screenshot.
+
+GOAL: Extract **every legible piece of text**, even partially visible, faded, or cropped. Structure your output for clarity. Do not guess, but always report what is visible.
+
+📄 FORMAT USING THESE CATEGORIES:
+
+---
+
+1. PAGE STRUCTURE
+- Page titles
+- Interface headers or section labels
+- Navigation bars or visible URLs
+
+2. IDENTIFIERS & DATA
+- Operator or user names
+- Sample IDs, test references
+- Materials, dates, batch numbers
+
+3. INTERFACE ELEMENTS (MANDATORY SCAN)
+- Button labels (e.g., RAZ, SAVE)
+- Tabs (e.g., MATERIAL, OBSERVATIONS)
+- Sidebars, form field labels
+
+4. SYSTEM MESSAGES
+- Connection or server errors
+- Domains, IP addresses, server notices
+
+5. METADATA
+- Standard references (e.g., "NF EN ####-#")
+- Version numbers, document codes, timestamps
+
+6. UNCLEAR / CROPPED TEXT
+- Logos, partial lines (use “[...]” for truncated)
+- Background/faded elements, labels not fully legible
+
+---
+
+RULES:
+- Preserve punctuation, case, accents exactly.
+- Include duplicates if text appears more than once.
+- Never skip faint or partial text; use “[...]” if incomplete.
+- Even if cropped, report as much as possible from any UI region.
+
+This prompt is designed to generalize across all web portals, technical forms, or reports. Prioritize completeness over certainty. Do not ignore UI components or system messages.
+
+Résultats:
+🔵 Paramètres actifs LLM:
+{
+  "temperature": 1.3,
+  "top_p": 0.85,
+  "presence_penalty": 0.1,
+  "frequency_penalty": 0.15,
+  "stop": [],
+  "stream": false,
+  "n": 1,
+  "seed": 0,
+  "mirostat": 0,
+  "mirostat_eta": 0.1,
+  "mirostat_tau": 5.0,
+  "top_k": 35,
+  "min_p": 0.06,
+  "repeat_penalty": 1.15,
+  "repeat_last_n": 128,
+  "tfs_z": 1.0,
+  "num_keep": 0,
+  "num_predict": 2048,
+  "num_ctx": 16384,
+  "num_batch": 2048
+}
+ AgentVisionOCR: Extraction OCR sur image_145435.png
+
+🔵 Résultat pour l'image output/ticket_T11143/T11143_20250422_084617/attachments/image_145435.png:
+Texte extrait:
+### **Page Structure:**
+
+*   Page title: "Essai au bleu de méthylène (MB) - NF EN 933-9 (02-2022)"
+*   Interface header: "RG-LAB"
+*   Navigation bar/visible URL: Not visible
+*   Sidebars/form field labels:
+    *   "MATERIEL"
+    *   "OBSERVATIONS"
+
+### **Identifiers & Data:** 
+
+No legible identifiers/data present in the image.
+
+### **Interface Elements:** 
+
+*   Button labels: None fully visible. One partially cropped button appears to start with an ellipsis "...".
+
+### **System Messages** 
+
+None are apparent from the interface elements shown, although partial text could suggest a server message or error code ("[...]", "[...]").
+
+### **Metadata**
+
+*   Standard references: NF EN 933-9 (02-2022)
+
+### **Unclear/Cropped Text**:
+
+The lower section contains a faded URL and some metadata fields that appear not to be filled out or have been intentionally hidden for privacy/security reasons ("[...]"). The top left corner shows part of what might be another standard reference or version number ("RG-LAB") but is too cropped to interpret clearly.
+
+Nouveau prompt: