fix(rag+llm): 60 wide treffer, Pfad-Dedup, 100k tool payload

(Nachtrag: vorheriger Commit enthielt nur telegram_bot.)

homelab-ai-bot/llm.py

@@ -450,7 +450,7 @@ def ask_with_tools(question: str, tool_handlers: dict, session_id: str = None, d
                         question
                         + " Versicherung Beitrag Beitragsrechnung Jahresbetrag"
                     )
-                _rag_res = _rag_fn(query=_rag_q, top_k=25)
+                _rag_res = _rag_fn(query=_rag_q, top_k=60)
                 if _rag_res and not _rag_res.startswith("Keine"):
                     log.info("RAG-Pflicht: %d Zeichen — loesche Session-History", len(str(_rag_res)))
                     messages = [
@@ -467,9 +467,9 @@ def ask_with_tools(question: str, tool_handlers: dict, session_id: str = None, d
                         {"role": "assistant", "content": None,
                             "tool_calls": [{"id": "forced_rag", "type": "function",
                                 "function": {"name": "rag_search",
-                                    "arguments": json.dumps({"query": _rag_q, "top_k": 25})}}]},
+                                    "arguments": json.dumps({"query": _rag_q, "top_k": 60})}}]},
                         {"role": "tool", "tool_call_id": "forced_rag",
-                            "content": str(_rag_res)[:32000]},
+                            "content": str(_rag_res)[:100000]},
                         {"role": "user", "content": question},
                     ]
             except Exception as e:

homelab-ai-bot/tools/rag.py

@@ -21,10 +21,10 @@ OLLAMA_EMBED_URL = "http://100.84.255.83:11434/api/embeddings"
 EMBED_MODEL = "nomic-embed-text"
 
 MIN_TOP_K = 5
-# Breite Übersichten: mehr ES-Runden, mehr distinct Treffer
+# Breite Übersichten: mehr ES-Runden, mehr distinct Treffer (pro vollem Pfad docnm_kwd)
-MAX_TOP_K_NORMAL = 15
+MAX_TOP_K_NORMAL = 25
-MAX_TOP_K_WIDE = 25
+MAX_TOP_K_WIDE = 60
-ES_SIZE_CAP = 150
+ES_SIZE_CAP = 200
 
 TOOLS = [
     {
@@ -146,6 +146,11 @@ def _dedup_key(name: str) -> str:
     return key.lower()
 
 
+def _dedup_key_full_doc(name: str) -> str:
+    """Ein Chunk pro vollem docnm_kwd — gleicher Dateiname in verschiedenen Ordnern bleibt getrennt."""
+    return re.sub(r"\s+", " ", (name or "").strip().lower())
+
+
 def _es_hybrid_search(query: str, es_size: int) -> dict:
     qvec = _ollama_embed(query)
     if not qvec:
@@ -273,16 +278,25 @@ _WIDE_SUBQUERIES = [
 ]
 
 
-def _merge_hits_from_queries(queries: list[str], es_size: int, pool_cap: int) -> tuple[list, str | None]:
+def _merge_hits_from_queries(
-    """Führt mehrere Hybrid-Suchen aus; pro Dokument höchster Score."""
+    queries: list[str],
+    es_size: int,
+    pool_cap: int,
+    *,
+    full_path_dedup: bool = False,
+) -> tuple[list, str | None]:
+    """Führt mehrere Hybrid-Suchen aus; pro Dedup-Key höchster Score."""
     best: dict[str, dict] = {}
     last_err: str | None = None
 
+    def dkey(dn: str) -> str:
+        return _dedup_key_full_doc(dn) if full_path_dedup else _dedup_key(dn)
+
     def absorb(hits: list) -> None:
         for h in hits:
             src = h.get("_source") or {}
             dn = src.get("docnm_kwd") or "?"
-            dk = _dedup_key(dn)
+            dk = dkey(dn)
             sc = float(h.get("_score") or 0.0)
             old = best.get(dk)
             if old is None or sc > float(old.get("_score") or 0.0):
@@ -318,15 +332,20 @@ def handle_rag_search(query: str, top_k: int = 8, **kw):
         for sq in _WIDE_SUBQUERIES:
             if sq.lower() not in qstrip.lower():
                 subqs.append(sq)
-        pool_cap = max(top_k * 3, 45)
+        pool_cap = max(top_k * 5, 120)
-        hits, err = _merge_hits_from_queries(subqs[:16], es_size, pool_cap=pool_cap)
+        hits, err = _merge_hits_from_queries(
+            subqs[:22],
+            es_size,
+            pool_cap=pool_cap,
+            full_path_dedup=True,
+        )
         if err and not hits:
             return f"Fehler bei der Dokumentensuche: {err}"
         header = (
-            f"**Breitensuche ({len(subqs[:16])} Anfragen gemerged) fuer '{qstrip}' — "
+            f"**Breitensuche ({len(subqs[:22])} Anfragen, Dedup=voller Pfad) '{qstrip}' — "
-            f"{len(hits)} Kandidaten, zeige bis {top_k} distinct:**\n"
+            f"{len(hits)} Kandidaten, zeige bis {top_k}:**\n"
         )
-        snip_len = 750
+        snip_len = 400
     else:
         data = _es_hybrid_search(qstrip, es_size)
         if "_error" in data:
@@ -342,12 +361,15 @@ def handle_rag_search(query: str, top_k: int = 8, **kw):
     lines: list[str] = []
     count = 0
 
+    def out_dkey(doc_name: str) -> str:
+        return _dedup_key_full_doc(doc_name) if wide else _dedup_key(doc_name)
+
     for h in hits:
         if count >= top_k:
             break
         src = h.get("_source") or {}
         doc_name = src.get("docnm_kwd") or "?"
-        dk = _dedup_key(doc_name)
+        dk = out_dkey(doc_name)
         if dk in seen_docs:
             continue
         seen_docs.add(dk)