RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.
Go agent-harnessagent-native-databaseagentic-aiagentic-retrievalagentic-searchaiai-agentscontext-enginecontext-engineeringcontext-managementharness-engineeringknowledge-compilationragretrieval-augmented-generationretrieval-harnesssearch-harness
fix: handle empty final tool answers (#19872)
Mage Lu7b1b67ae4d62
diff --git a/rag/llm/chat_model.py b/rag/llm/chat_model.pyindex 684ac7fcde..860c1527e3 100644--- a/rag/llm/chat_model.py+++ b/rag/llm/chat_model.py@@ -653,11 +653,19 @@ class Base(ABC):raise Exception(f"500 response structure error. Response: {response}")if not hasattr(response.choices[0].message, "tool_calls") or not response.choices[0].message.tool_calls:+ content = response.choices[0].message.content+ if not content or not content.strip():+ if not ans.strip():+ return (+ (f"{ERROR_PREFIX}: {LLMErrorCode.ERROR_MODEL} - Model returned an empty final answer without tool calls (finish_reason={response.choices[0].finish_reason})."),+ tk_count,+ )+ content = ""_reasoning = getattr(response.choices[0].message, "reasoning_content", None) or getattr(response.choices[0].message, "reasoning", None)if _reasoning:ans += "<think>" + _reasoning + "</think>"- ans += response.choices[0].message.content+ ans += contentif response.choices[0].finish_reason == "length":ans = self._length_stop(ans)diff --git a/test/unit_test/rag/llm/test_empty_tool_response.py b/test/unit_test/rag/llm/test_empty_tool_response.pynew file mode 100644index 0000000000..1c45b4eda9--- /dev/null+++ b/test/unit_test/rag/llm/test_empty_tool_response.py@@ -0,0 +1,89 @@+import asyncio+from types import SimpleNamespace++import httpx+import pytest+from openai import AsyncOpenAI++from rag.llm.chat_model import Base++[email protected]("content", [None, "", " "])[email protected]("finish_reason", ["stop", "length"])[email protected]("max_retries", [0, 2])+def test_empty_final_answer_is_explicit_model_error(content, finish_reason, max_retries):+ answer, tokens, requests = asyncio.run(_run_tool_rounds(content, finish_reason, max_retries=max_retries))+ assert answer.startswith("**ERROR**: MODEL_ERROR")+ assert "empty" in answer.lower()+ assert finish_reason in answer+ assert tokens == 6+ assert len(requests) == 2++[email protected]("reasoning_field", ["reasoning_content", "reasoning"])+def test_tool_call_with_text_preserves_reasoning_and_length_notice(reasoning_field):+ answer, tokens, requests = asyncio.run(_run_tool_rounds("Found a document.", "length", reasoning_field=reasoning_field, reasoning="Reasoning."))+ assert answer == "<think>Reasoning.</think>Found a document....\nThe answer is truncated by your chosen LLM due to its limitation on context length."+ assert "**ERROR**" not in answer+ assert tokens == 6+ assert len(requests) == 2++[email protected]("content", [None, "", " "])+def test_empty_final_answer_preserves_accumulated_verbose_tool_output(content):+ answer, tokens, requests = asyncio.run(+ _run_tool_rounds(+ content,+ "stop",+ reasoning_field="reasoning_content",+ reasoning="Reasoning.",+ verbose_tool_use=True,+ )+ )++ assert answer == ('<tool_call>{\n "name": "search",\n "args": {},\n "result": "Synthetic document."\n}</tool_call><think>Reasoning.</think>')+ assert tokens == 6+ assert len(requests) == 2+++async def _run_tool_rounds(content, finish_reason, *, max_retries=0, reasoning_field=None, reasoning=None, verbose_tool_use=False):+ requests = []++ def respond(request):+ requests.append(request)+ if len(requests) == 1:+ message = {"role": "assistant", "content": None, "tool_calls": [{"id": "call-test", "type": "function", "function": {"name": "search", "arguments": "{}"}}]}+ reason = "tool_calls"+ else:+ message = {"role": "assistant", "content": content}+ if reasoning_field:+ message[reasoning_field] = reasoning+ reason = finish_reason+ return httpx.Response(+ 200,+ json={+ "id": "test-completion",+ "object": "chat.completion",+ "created": 0,+ "model": "test-model",+ "choices": [{"index": 0, "message": message, "finish_reason": reason}],+ "usage": {"prompt_tokens": 2, "completion_tokens": 1, "total_tokens": 3},+ },+ )++ async def search(name, arguments):+ assert name == "search" and arguments == {}+ return "Synthetic document."++ model = Base.__new__(Base)+ model.model_name = "test-model"+ model.max_retries = max_retries+ model.max_rounds = 5+ model.tools = [{"type": "function", "function": {"name": "search", "parameters": {"type": "object"}}}]+ model.toolcall_session = SimpleNamespace(tool_call_async=search)+ if not verbose_tool_use:+ model._verbose_tool_use = lambda *_args: ""+ async with AsyncOpenAI(api_key="offline-placeholder", base_url="https://offline.invalid/v1", http_client=httpx.AsyncClient(transport=httpx.MockTransport(respond), trust_env=False)) as client:+ model.async_client = client+ answer, tokens = await model.async_chat_with_tools("Use the supplied document.", [{"role": "user", "content": "Find it."}])+ return answer, tokens, requests