orbitalo/homelab-brain
1
0
Fork 0
Code Issues 37 Pull requests Projects Releases Packages Wiki Activity Actions

fix(rag): strip session history when forced RAG fires

Browse source 
Poisoned session history (4x wrong answers) overrode RAG results.
Now: when doc keywords trigger forced rag_search, rebuild messages
from scratch with ONLY system prompt + RAG results + question.
No session history = no poisoning.
This commit is contained in:
Homelab Cursor 2026-03-26 16:51:29 +01:00
parent 6abafbc792
commit 7b6543be90
1 changed files with 15 additions and 15 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

28
homelab-ai-bot/llm.py
View file

@ -412,7 +412,7 @@ def ask_with_tools(question: str, tool_handlers: dict, session_id: str = None) -
messages.append({"role": "user", "content": question}) messages.append({"role": "user", "content": question})
# --- RAG-Pflicht: Bei Doc-Keywords rag_search DIREKT aufrufen (nicht LLM) --- # --- RAG-Pflicht: Bei Doc-Keywords rag_search DIREKT aufrufen ---
_DOC_KW = [ _DOC_KW = [
"versicherung", "vertrag", "vertraege", "dokument", "rente", "versicherung", "vertrag", "vertraege", "dokument", "rente",
"finanzamt", "steuer", "grundsteuer", "familienbuch", "urkunde", "finanzamt", "steuer", "grundsteuer", "familienbuch", "urkunde",
@ -429,20 +429,20 @@ def ask_with_tools(question: str, tool_handlers: dict, session_id: str = None) -
log.info("RAG-Pflicht: forciere rag_search fuer: %s", question[:80]) log.info("RAG-Pflicht: forciere rag_search fuer: %s", question[:80])
_rag_res = _rag_fn(query=question, top_k=8) _rag_res = _rag_fn(query=question, top_k=8)
if _rag_res and not _rag_res.startswith("Keine"): if _rag_res and not _rag_res.startswith("Keine"):
_fake_tc = [{"id": "forced_rag", "type": "function", log.info("RAG-Pflicht: %d Zeichen — loesche Session-History", len(str(_rag_res)))
messages = [
{"role": "system", "content": _full_prompt
+ "\n\nWICHTIG: Ignoriere fruehere Antworten. "
+ "Die Dokumentensuche unten ist die einzige Wahrheit. "
+ "Beantworte die Frage NUR basierend auf diesen Suchergebnissen."},
{"role": "assistant", "content": None,
"tool_calls": [{"id": "forced_rag", "type": "function",
"function": {"name": "rag_search", "function": {"name": "rag_search",
"arguments": json.dumps({"query": question, "top_k": 8})}}] "arguments": json.dumps({"query": question, "top_k": 8})}}]},
messages.insert(-1, {"role": "assistant", "content": None, {"role": "tool", "tool_call_id": "forced_rag",
"tool_calls": _fake_tc}) "content": str(_rag_res)[:3000]},
messages.insert(-1, {"role": "tool", "tool_call_id": "forced_rag", {"role": "user", "content": question},
"content": str(_rag_res)[:3000]}) ]
_suffix = (
"\n\n[Oben siehst du die Ergebnisse der Dokumentensuche. "
"Beantworte die Frage NUR basierend auf diesen Ergebnissen. "
"Erfinde nichts dazu.]"
)
messages[-1] = {"role": "user", "content": question + _suffix}
log.info("RAG-Pflicht: %d Zeichen injiziert", len(str(_rag_res)))
except Exception as e: except Exception as e:
log.warning("RAG-Pflicht Fehler: %s", e) log.warning("RAG-Pflicht Fehler: %s", e)