diff --git a/rag.py b/rag.py
index db6a9b8..2ecac8f 100644
--- a/rag.py
+++ b/rag.py
@@ -1,58 +1,107 @@
 #!/usr/bin/env python3
-import faiss, json, requests, readline, numpy as np
-from rich import print
+"""
+Mini RAG interactif :
+• Recherche sémantique FAISS sur le corpus.idx / corpus.meta.json
+• Contexte (top‑k=4 passages) envoyé à Mistral‑7B via Ollama.
+
+Robuste aux différentes sorties de BGEM3FlagModel.encode.
+"""
+import json
+import readline
+from pathlib import Path
+
+import faiss
+import numpy as np
+import requests
 from FlagEmbedding import BGEM3FlagModel
+from rich import print
 
-# --- chargements -------------------------------------------------------------
-idx   = faiss.read_index("corpus.idx")
-meta  = json.load(open("corpus.meta.json"))
-model = BGEM3FlagModel("BAAI/bge-m3", device="cpu")      # même qu’à l’indexation
+# ---------------------------------------------------------------------------
+# Chargements initiaux
+# ---------------------------------------------------------------------------
+IDX_FILE = Path("corpus.idx")
+META_FILE = Path("corpus.meta.json")
 
-# simple aide mémoire pour retrouver rapidement un passage
-def fetch_passage(i):
+if not IDX_FILE.exists() or not META_FILE.exists():
+    raise SystemExit("[bold red]Erreur :[/] index absent. Lancez d'abord index.py !")
+
+index = faiss.read_index(str(IDX_FILE))
+meta = json.loads(META_FILE.read_text())
+model = BGEM3FlagModel("BAAI/bge-m3", device="cpu")
+
+# ---------------------------------------------------------------------------
+# Utilitaires
+# ---------------------------------------------------------------------------
+
+def _normalize(x: np.ndarray) -> np.ndarray:
+    """L2‑normalize each row (tokens=float32)."""
+    return x / (np.linalg.norm(x, axis=1, keepdims=True) + 1e-12)
+
+
+def embed(texts):
+    """Encode list[str] → ndarray (n, dim), quelle que soit la sortie lib."""
+    out = model.encode(texts)
+    # Possible shapes :
+    #  • ndarray
+    #  • dict {"embedding": ndarray} ou {"embeddings": ndarray}
+    #  • dict {"sentence_embeds": [...]} etc.
+    if isinstance(out, np.ndarray):
+        arr = out
+    elif isinstance(out, dict):
+        # pick the first ndarray-like value
+        for v in out.values():
+            if isinstance(v, (list, tuple)) or hasattr(v, "shape"):
+                arr = np.asarray(v)
+                break
+        else:
+            raise TypeError("encode() dict sans clé embedding !")
+    else:  # list[list[float]] etc.
+        arr = np.asarray(out)
+    return _normalize(arr.astype("float32"))
+
+
+def fetch_passage(i: int) -> str:
     m = meta[i]
     return f"[{m['file']} · part {m['part']}] {m['text']}"
 
-def ask_llm(prompt):
-    r = requests.post("http://127.0.0.1:11434/api/generate", json={
-        "model": "mistral7b-fast",
-        "prompt": prompt,
-        "stream": False,
-        "options": {"temperature":0.2, "num_predict":512}
-    }, timeout=300)
+
+def ask_llm(prompt: str) -> str:
+    r = requests.post(
+        "http://127.0.0.1:11434/api/generate",
+        json={
+            "model": "mistral7b-fast",
+            "prompt": prompt,
+            "stream": False,
+            "options": {"temperature": 0.2, "num_predict": 512},
+        },
+        timeout=300,
+    )
     return r.json()["response"]
 
-# --- boucle interactive ------------------------------------------------------
+# ---------------------------------------------------------------------------
+# Boucle interactive
+# ---------------------------------------------------------------------------
+print("[bold green]RAG prêt.[/] Posez vos questions ! (Ctrl‑D pour sortir)")
 while True:
     try:
         q = input("❓ > ").strip()
-        if not q: continue
-    except (KeyboardInterrupt, EOFError):
-        print("\nBye."); break
+        if not q:
+            continue
+    except (EOFError, KeyboardInterrupt):
+        print("\n[dim]Bye.[/]")
+        break
 
-    # embeddings & recherche FAISS (top-k=4)
-    # remplace ces deux lignes (32-34)
-    # q_emb = model.encode([q], normalize_embeddings=True)
-    # D, I  = idx.search(q_emb.astype("float32"), 4)
+    q_emb = embed([q])  # (1, dim)
+    D, I = index.search(q_emb, 4)
+    ctx = "\n\n".join(fetch_passage(int(idx)) for idx in I[0])
 
-    emb = model.encode([q])           # ndarray (1, 1024)
-    if isinstance(emb, dict):         # selon la version de FlagEmbedding
-        emb = emb.get("embedding") or emb.get("embeddings")
-    q_emb = emb[0] / np.linalg.norm(emb[0])   # L2 normalisation
+    prompt = (
+        "<system>Réponds en français, précis et factuel.</system>\n"
+        f"<context>{ctx}</context>\n"
+        f"<user>{q}</user>"
+    )
 
-    D, I = idx.search(q_emb.astype("float32").reshape(1, -1), 4)
-
-
-    ctx_blocks = []
-    for rank, idx_id in enumerate(I[0]):
-        ctx_blocks.append(fetch_passage(idx_id))
-    context = "\n\n".join(ctx_blocks)
-
-    prompt = f"""<system>Réponds en français, précis et factuel.</system>
-<context>{context}</context>
-<user>{q}</user>"""
-
-    print("\n[bold]Réponse :[/]\n")
+    print("\n[bold]Réponse :[/]\n")
     print(ask_llm(prompt))
     print("\n[dim]--- contexte utilisé ---[/]")
-    print(context)
+    print(ctx)