RomanNum3ral
/
MicroFish
mirror of https://github.com/666ghj/MiroFish.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix(report): strip fabricated tool_result blocks to prevent LLM hallucination loop

This commit is contained in:
Ubuntu 2026-04-26 15:06:24 +00:00
parent 866ed421e2
commit 49f4da51b1
1 changed files with 15 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
backend/app/services/report_agent.py
View File

@ -1132,7 +1132,14 @@ class ReportAgent:
data["parameters"] = data.pop("params")
return True
return False
@staticmethod
def _strip_fake_tool_results(response: str) -> str:
"""Strip <tool_result> blocks fabricated by the LLM to prevent hallucination loops."""
cleaned = re.sub(r'<tool_result>.*?</tool_result>', '', response, flags=re.DOTALL)
cleaned = re.sub(r'\n{3,}', '\n\n', cleaned)
return cleaned.strip()
def _get_tools_description(self) -> str:
"""Generate tool description text"""
desc_parts = ["Available tools:"]
@ -1322,7 +1329,7 @@ class ReportAgent:
logger.warning(t('report.sectionIterNone', title=section.title, iteration=iteration + 1))
# If iterations remain, append a message and retry
if iteration < max_iterations - 1:
messages.append({"role": "assistant", "content": "(empty response)"})
messages.append({"role": "assistant", "content": ReportAgent._strip_fake_tool_results("(empty response)")})
messages.append({"role": "user", "content": "Please continue generating content."})
continue
# Last iteration also returned None; break out of the loop and force a conclusion
@ -1344,7 +1351,7 @@ class ReportAgent:
if conflict_retries <= 2:
# First two occurrences: discard this response and ask the LLM to reply again
messages.append({"role": "assistant", "content": response})
messages.append({"role": "assistant", "content": ReportAgent._strip_fake_tool_results(response)})
messages.append({
"role": "user",
"content": (
@ -1384,7 +1391,7 @@ class ReportAgent:
if has_final_answer:
# Insufficient tool calls — reject and require more tool usage
if tool_calls_count < min_tool_calls:
messages.append({"role": "assistant", "content": response})
messages.append({"role": "assistant", "content": ReportAgent._strip_fake_tool_results(response)})
unused_tools = all_tools - used_tools
unused_hint = f"(These tools have not been used yet, consider trying them: {', '.join(unused_tools)})" if unused_tools else ""
messages.append({
@ -1414,7 +1421,7 @@ class ReportAgent:
if has_tool_calls:
# Tool quota exhausted → notify clearly and require Final Answer output
if tool_calls_count >= self.MAX_TOOL_CALLS_PER_SECTION:
messages.append({"role": "assistant", "content": response})
messages.append({"role": "assistant", "content": ReportAgent._strip_fake_tool_results(response)})
messages.append({
"role": "user",
"content": REACT_TOOL_LIMIT_MSG.format(
@ -1462,7 +1469,7 @@ class ReportAgent:
if unused_tools and tool_calls_count < self.MAX_TOOL_CALLS_PER_SECTION:
unused_hint = REACT_UNUSED_TOOLS_HINT.format(unused_list=", ".join(unused_tools))
messages.append({"role": "assistant", "content": response})
messages.append({"role": "assistant", "content": ReportAgent._strip_fake_tool_results(response)})
messages.append({
"role": "user",
"content": REACT_OBSERVATION_TEMPLATE.format(
@ -1477,7 +1484,7 @@ class ReportAgent:
continue
# ── Case 3: Neither a tool call nor a Final Answer ──
messages.append({"role": "assistant", "content": response})
messages.append({"role": "assistant", "content": ReportAgent._strip_fake_tool_results(response)})
if tool_calls_count < min_tool_calls:
# Insufficient tool calls — recommend unused tools
@ -1867,7 +1874,7 @@ class ReportAgent:
tool_calls_made.append(call)
# Append the results to the messages
messages.append({"role": "assistant", "content": response})
messages.append({"role": "assistant", "content": ReportAgent._strip_fake_tool_results(response)})
observation = "\n".join([f"[{r['tool']} result]\n{r['result']}" for r in tool_results])
messages.append({
"role": "user",