Skip to content

C
coderai
Commit 886ea8f4 authored by Stefy Lanza (nextime / spora )'s avatar Stefy Lanza (nextime / spora )
Browse files
Options

Fix tool parsing: deduplicate tool calls, strip raw format from streaming content 

- Add seen_signatures set to extract_tool_calls() to prevent duplicates
- Add strip_tool_calls_from_content() method to remove <tool>...</tool> tags
- Filter tool format from each chunk in real-time during streaming
- Simplify post-stream tool call handling since content is already cleaned
- Also handle non-streaming responses for tool call content cleanup
parent 821e40dd
Hide whitespace changes
Inline Side-by-side
Showing with 89 additions and 42 deletions
coderai
View file @ 886ea8f4
......@@ -282,6 +282,7 @@ class ToolCallParser:
text = self._filter_malformed_content(text)
tool_calls = []
seen_signatures = set() # Track seen tool calls to avoid duplicates
# Look for function calls in various formats
# Format 1: <tool> or <function> tags with JSON content
......@@ -293,14 +294,18 @@ class ToolCallParser:
try:
tool_data = json.loads(match.strip())
if 'name' in tool_data and 'arguments' in tool_data:
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:16]}",
"type": "function",
"function": {
"name": tool_data["name"],
"arguments": json.dumps(tool_data["arguments"])
}
})
# Create a signature to detect duplicates
sig = (tool_data['name'], json.dumps(tool_data['arguments'], sort_keys=True))
if sig not in seen_signatures:
seen_signatures.add(sig)
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:16]}",
"type": "function",
"function": {
"name": tool_data["name"],
"arguments": json.dumps(tool_data["arguments"])
}
})
continue
except json.JSONDecodeError:
pass
......@@ -309,14 +314,18 @@ class ToolCallParser:
try:
tool_data = self._parse_nested_xml_tool(match)
if tool_data and 'name' in tool_data and 'arguments' in tool_data:
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:16]}",
"type": "function",
"function": {
"name": tool_data["name"],
"arguments": json.dumps(tool_data["arguments"]) if isinstance(tool_data["arguments"], dict) else str(tool_data["arguments"])
}
})
args_str = json.dumps(tool_data["arguments"]) if isinstance(tool_data["arguments"], dict) else str(tool_data["arguments"])
sig = (tool_data['name'], args_str)
if sig not in seen_signatures:
seen_signatures.add(sig)
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:16]}",
"type": "function",
"function": {
"name": tool_data["name"],
"arguments": args_str
}
})
except Exception:
pass
......@@ -328,14 +337,17 @@ class ToolCallParser:
data = json.loads(json_match.group())
if "function_call" in data:
fc = data["function_call"]
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:16]}",
"type": "function",
"function": {
"name": fc.get("name", ""),
"arguments": fc.get("arguments", "{}")
}
})
sig = (fc.get("name", ""), fc.get("arguments", "{}"))
if sig not in seen_signatures:
seen_signatures.add(sig)
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:16]}",
"type": "function",
"function": {
"name": fc.get("name", ""),
"arguments": fc.get("arguments", "{}")
}
})
except (json.JSONDecodeError, AttributeError):
pass
......@@ -347,27 +359,52 @@ class ToolCallParser:
for match in matches:
try:
args = json.loads(match.strip())
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:16]}",
"type": "function",
"function": {
"name": tool_name,
"arguments": json.dumps(args)
}
})
sig = (tool_name, json.dumps(args, sort_keys=True))
if sig not in seen_signatures:
seen_signatures.add(sig)
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:16]}",
"type": "function",
"function": {
"name": tool_name,
"arguments": json.dumps(args)
}
})
except json.JSONDecodeError:
# Try to use the raw text as arguments
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:16]}",
"type": "function",
"function": {
"name": tool_name,
"arguments": match.strip()
}
})
sig = (tool_name, match.strip())
if sig not in seen_signatures:
seen_signatures.add(sig)
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:16]}",
"type": "function",
"function": {
"name": tool_name,
"arguments": match.strip()
}
})
return tool_calls if tool_calls else None
def strip_tool_calls_from_content(self, text: str) -> str:
"""Remove tool call format from text after extracting tool calls."""
if not text:
return text
# Remove <tool>...</tool> and <function>...</function> patterns
text = re.sub(r'<tool>.*?</tool>', '', text, flags=re.DOTALL)
text = re.sub(r'<function>.*?</function>', '', text, flags=re.DOTALL)
# Also remove common tool name tags like <read>...</read>, <exec>...</exec>
for tool_name in ['read', 'write', 'exec', 'browser', 'message', 'web_search', 'web_fetch',
'memory_search', 'memory_get', 'sessions_list', 'sessions_send', 'tts', 'canvas', 'nodes']:
text = re.sub(rf'<{tool_name}>.*?</{tool_name}>', '', text, flags=re.DOTALL)
# Clean up excessive newlines left from removal
text = re.sub(r'\n{3,}', '\n\n', text)
return text.strip()
def format_tools_for_prompt(tools: List[Tool], messages: List[ChatMessage]) -> List[ChatMessage]:
"""Format tools into the system message or add a tool description."""
......@@ -1391,6 +1428,7 @@ class VulkanBackend(ModelBackend):
total_content = ""
chunk_count = 0
has_tools = tools is not None # Track if tools are available
# Check if we should use manual formatting based on detected template
# Always use manual formatting when tools are present, since Jinja templates often fail with tool messages
......@@ -1950,6 +1988,11 @@ async def stream_chat_response(
chunk_count += 1
# Filter malformed content from each chunk
filtered_chunk = filter_malformed_content(chunk)
# Also filter out tool call format from content when tools are available
if tools:
filtered_chunk = model_manager.tool_parser.strip_tool_calls_from_content(filtered_chunk)
if not filtered_chunk:
print(f"DEBUG: filtered_chunk was empty (original chunk: {repr(chunk[:50])})")
continue
......@@ -1975,7 +2018,7 @@ async def stream_chat_response(
if not generated_text.strip():
print(f"DEBUG: Warning - no content generated!")
# Check for tool calls in complete output
# Check for tool calls in complete output (for API response format)
if tools:
# Convert tools back to Tool objects for parsing
from typing import cast
......@@ -1989,6 +2032,8 @@ async def stream_chat_response(
tool_objects.append(Tool(type=t.get("type", "function"), function=tool_func))
tool_calls = model_manager.tool_parser.extract_tool_calls(generated_text, tool_objects)
if tool_calls:
# Tool calls were extracted and stripped from content during streaming
# Just send the tool_calls chunk
data = {
"id": completion_id,
"object": "chat.completion.chunk",
......@@ -2071,8 +2116,10 @@ async def generate_chat_response(
tool_objects.append(Tool(type=t.get("type", "function"), function=tool_func))
tool_calls = model_manager.tool_parser.extract_tool_calls(generated_text, tool_objects)
if tool_calls:
# Strip tool call format from content so user doesn't see raw tags
clean_content = model_manager.tool_parser.strip_tool_calls_from_content(generated_text)
response_message["content"] = clean_content if clean_content.strip() else None
response_message["tool_calls"] = tool_calls
response_message["content"] = None
finish_reason = "tool_calls"
# Calculate token counts - rough estimate since we don't have direct access to tokenizer
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment