feat(extraction): appels LLM avancés côté worker (Ollama)

services/worker/pipelines/extract.py

@@ -6,6 +6,7 @@ import os
 import logging
 import re
 from typing import Dict, Any, List
+from services.worker.utils.llm_client import WorkerLLMClient
 
 logger = logging.getLogger(__name__)
 
@@ -20,6 +21,13 @@ def run(doc_id: str, ctx: Dict[str, Any]) -> None:
         # Extraction basique
         entities = _extract_basic_entities(ocr_text, document_type)
 
+        # Extraction avancée via LLM (merge non destructif)
+        llm = WorkerLLMClient()
+        prompt = _build_extraction_prompt(ocr_text[:3000] if ocr_text else "", document_type)
+        llm_response = llm.generate(prompt)
+        llm_json = WorkerLLMClient.extract_first_json(llm_response) or {}
+        entities = _merge_entities_basic_with_llm(entities, llm_json)
+
         ctx.update({
             "extracted_entities": entities,
             "entities_count": len(entities)
@@ -63,4 +71,33 @@ def _extract_basic_entities(text: str, doc_type: str) -> List[Dict[str, Any]]:
             "confidence": 0.8
         })
 
-    return entities
+    return entities
+
+
+def _build_extraction_prompt(text: str, doc_type: str) -> str:
+    return f"""
+Tu es un extracteur d'entités pour documents notariaux.
+Type de document: {doc_type}
+Extrait en JSON strict les objets: identites, adresses, biens, entreprises, montants, dates.
+Réponds UNIQUEMENT par un JSON.
+
+TEXTE:
+{text}
+"""
+
+
+def _merge_entities_basic_with_llm(basic: List[Dict[str, Any]], advanced: Dict[str, Any]) -> List[Dict[str, Any]]:
+    merged = list(basic)
+    if not isinstance(advanced, dict):
+        return merged
+    # Aplatit les entités LLM en liste simple type/value pour compatibilité minimale
+    for key in ["identites", "adresses", "biens", "entreprises", "montants", "dates"]:
+        items = advanced.get(key, []) or []
+        for item in items:
+            try:
+                value = item.get("adresse_complete") or item.get("date") or item.get("montant") or item.get("nom") or item.get("description") or str(item)
+                if value:
+                    merged.append({"type": key, "value": value, "confidence": item.get("confidence", 0.8)})
+            except Exception:
+                continue
+    return merged

services/worker/utils/llm_client.py

@@ -0,0 +1,58 @@
+"""
+Client LLM minimal côté worker pour appels à Ollama
+"""
+
+import os
+import json
+import logging
+from typing import Optional
+
+import requests
+
+logger = logging.getLogger(__name__)
+
+
+class WorkerLLMClient:
+    def __init__(self) -> None:
+        self.base_url = os.getenv("OLLAMA_BASE_URL", "http://ollama:11434")
+        self.default_model = os.getenv("OLLAMA_DEFAULT_MODEL", "llama3:8b")
+        self.session = requests.Session()
+
+    def generate(self, prompt: str, model: Optional[str] = None, max_tokens: int = 2000) -> str:
+        model_name = model or self.default_model
+        url = f"{self.base_url}/api/generate"
+        payload = {
+            "model": model_name,
+            "prompt": prompt,
+            "stream": False,
+            "options": {
+                "temperature": 0.1,
+                "top_p": 0.9,
+                "max_tokens": max_tokens,
+            },
+        }
+        try:
+            resp = self.session.post(url, json=payload, timeout=120)
+            if resp.status_code != 200:
+                logger.error("Erreur LLM %s: %s", resp.status_code, resp.text)
+                raise RuntimeError(f"LLM HTTP {resp.status_code}")
+            data = resp.json()
+            return data.get("response", "")
+        except Exception as exc:
+            logger.error("Erreur appel LLM: %s", exc)
+            raise
+
+    @staticmethod
+    def extract_first_json(text: str) -> Optional[dict]:
+        try:
+            import re
+
+            m = re.search(r"\{[\s\S]*\}", text)
+            if not m:
+                return None
+            return json.loads(m.group(0))
+        except Exception as exc:
+            logger.warning("JSON non parsé depuis la réponse LLM: %s", exc)
+            return None
+
+