diff --git a/homelab-ai-bot/tools/rag.py b/homelab-ai-bot/tools/rag.py
index 9fd62c10..b08ba3e3 100644
--- a/homelab-ai-bot/tools/rag.py
+++ b/homelab-ai-bot/tools/rag.py
@@ -6,7 +6,9 @@ RAGFlow bleibt Ingestion; Suche geht direkt an ES (Issue #51).
 import base64
 import json
 import logging
+import os
 import re
+import time
 import urllib.error
 import urllib.request
 
@@ -466,6 +468,103 @@ def _expand_multilingual(q: str) -> list:
     return variants[:8]
 # --- /Multi-Query-Erweiterung ---------------------------------------------
 
+
+# --- LLM Query Rewrite (Stufe B) ---
+_LLM_REWRITE_CACHE: dict = {}
+_LLM_REWRITE_TTL = 3600
+_LLM_REWRITE_MAX = 6
+_LLM_REWRITE_TIMEOUT = 8
+_LLM_REWRITE_MODEL = "gpt-4o-mini"
+_LLM_REWRITE_PROMPT = (
+    "Du bist ein Query-Rewriter fuer eine deutschsprachige Dokumenten-Suche "
+    "(Vertraege, Versicherungen, Rechnungen, Bescheide, Kaufvertraege, Kontoauszuege). "
+    "Formuliere die Nutzer-Frage in {n} unterschiedliche kompakte Such-Queries um. "
+    "Nutze Synonyme, Fachbegriffe, Zahl-/Kosten-/Summenbegriffe und sowohl deutsche "
+    "als auch englische Termini. Jede Query: 2-6 Woerter, Keywords statt Saetze, "
+    "keine Fragezeichen. Antworte NUR als JSON-Liste von Strings, nichts sonst."
+)
+
+
+def _openai_key() -> str:
+    try:
+        from core import config as _cfg  # type: ignore
+        v = (_cfg.parse_config().raw.get("OPENAI_API_KEY") or "").strip()
+        if v:
+            return v
+    except Exception:
+        pass
+    return (os.environ.get("OPENAI_API_KEY") or "").strip()
+
+
+def _llm_query_rewrite(query: str, n: int = 4) -> list:
+    """gpt-4o-mini -> bis zu n DE/EN-Reformulierungen. Cache + Fehler-Fallback."""
+    q = (query or "").strip()
+    if not q or n <= 0:
+        return []
+    cache_key = f"{n}::{q.lower()}"
+    now = time.time()
+    entry = _LLM_REWRITE_CACHE.get(cache_key)
+    if entry and (now - entry[0]) < _LLM_REWRITE_TTL:
+        return list(entry[1])
+    key = _openai_key()
+    if not key:
+        return []
+    body = json.dumps({
+        "model": _LLM_REWRITE_MODEL,
+        "messages": [
+            {"role": "system", "content": _LLM_REWRITE_PROMPT.format(n=n)},
+            {"role": "user", "content": q},
+        ],
+        "temperature": 0.3,
+        "max_tokens": 220,
+    }).encode()
+    req = urllib.request.Request(
+        "https://api.openai.com/v1/chat/completions",
+        data=body,
+        method="POST",
+        headers={
+            "Content-Type": "application/json",
+            "Authorization": f"Bearer {key}",
+        },
+    )
+    try:
+        with urllib.request.urlopen(req, timeout=_LLM_REWRITE_TIMEOUT) as resp:
+            data = json.load(resp)
+    except Exception as e:
+        log.warning("llm rewrite failed: %s", e)
+        return []
+    try:
+        txt = (data["choices"][0]["message"]["content"] or "").strip()
+    except Exception:
+        return []
+    m = re.search(r"\[[\s\S]*\]", txt)
+    raw = m.group(0) if m else txt
+    variants: list = []
+    try:
+        arr = json.loads(raw)
+        if isinstance(arr, list):
+            for x in arr:
+                if isinstance(x, str):
+                    s = x.strip()
+                    while s and s[-1] in ".?!;,":
+                        s = s[:-1]
+                    s = s.strip()
+                    if s and s.lower() != q.lower() and s not in variants:
+                        variants.append(s)
+    except Exception:
+        for line in txt.splitlines():
+            s = re.sub(r"^[\s\-\*\d\.\)]+", "", line).strip()
+            while s and s[-1] in ".?!;,":
+                s = s[:-1]
+            s = s.strip()
+            if s and s.lower() != q.lower() and s not in variants:
+                variants.append(s)
+    variants = variants[:n]
+    _LLM_REWRITE_CACHE[cache_key] = (now, variants)
+    log.info("llm rewrite %r -> %s variants", q[:60], len(variants))
+    return variants
+# --- /LLM Query Rewrite ---
+
 def handle_rag_search(query: str, top_k: int = 8, **kw):
     if not query or not query.strip():
         return "rag_search: query fehlt."
@@ -514,12 +613,39 @@ def handle_rag_search(query: str, top_k: int = 8, **kw):
         )
         snip_len = 500
     else:
-        data = _es_hybrid_search(qstrip, es_size)
-        if "_error" in data:
-            return f"Fehler bei der Dokumentensuche: {data['_error']}"
-        hits = (data.get("hits") or {}).get("hits") or []
-        header = f"**Dokumente fuer '{qstrip}' (bis {top_k}):**\n"
-        snip_len = 650
+        # Modus 3 (2026-04-16): LLM-Query-Rewrite via gpt-4o-mini.
+        # Fallback auf Single-Query bei API-Fehler / fehlendem Key.
+        rewrites = _llm_query_rewrite(qstrip, n=4)
+        if rewrites:
+            subqs = [qstrip] + [r for r in rewrites if r.lower() != qstrip.lower()]
+            subqs = subqs[:_LLM_REWRITE_MAX]
+            pool_cap = max(top_k * 5, 80)
+            hits, err = _merge_hits_from_queries(
+                subqs,
+                es_size,
+                pool_cap=pool_cap,
+                full_path_dedup=False,
+            )
+            if hits:
+                header = (
+                    f"**LLM-Rewrite ({len(subqs)} Varianten, gpt-4o-mini) '{qstrip}' — "
+                    f"{len(hits)} Kandidaten, zeige bis {top_k}:**\n"
+                )
+                snip_len = 600
+            else:
+                data = _es_hybrid_search(qstrip, es_size)
+                if "_error" in data:
+                    return f"Fehler bei der Dokumentensuche: {data['_error']}"
+                hits = (data.get("hits") or {}).get("hits") or []
+                header = f"**Dokumente fuer '{qstrip}' (bis {top_k}):**\n"
+                snip_len = 650
+        else:
+            data = _es_hybrid_search(qstrip, es_size)
+            if "_error" in data:
+                return f"Fehler bei der Dokumentensuche: {data['_error']}"
+            hits = (data.get("hits") or {}).get("hits") or []
+            header = f"**Dokumente fuer '{qstrip}' (bis {top_k}):**\n"
+            snip_len = 650
 
     if not hits:
         return f"Keine Ergebnisse fuer '{qstrip}' in der Wissensbasis gefunden."