feat: Implement Trillium notes executor for searching and creating notes via ETAPI

- Added `trillium.py` for searching and creating notes with Trillium's ETAPI.
- Implemented `search_notes` and `create_note` functions with appropriate error handling and validation.

feat: Add web search functionality using DuckDuckGo

- Introduced `web_search.py` for performing web searches without API keys.
- Implemented `search_web` function with result handling and validation.

feat: Create provider-agnostic function caller for iterative tool calling

- Developed `function_caller.py` to manage LLM interactions with tools.
- Implemented iterative calling logic with error handling and tool execution.

feat: Establish a tool registry for managing available tools

- Created `registry.py` to define and manage tool availability and execution.
- Integrated feature flags for enabling/disabling tools based on environment variables.

feat: Implement event streaming for tool calling processes

- Added `stream_events.py` to manage Server-Sent Events (SSE) for tool calling.
- Enabled real-time updates during tool execution for enhanced user experience.

test: Add tests for tool calling system components

- Created `test_tools.py` to validate functionality of code execution, web search, and tool registry.
- Implemented asynchronous tests to ensure proper execution and result handling.

chore: Add Dockerfile for sandbox environment setup

- Created `Dockerfile` to set up a Python environment with necessary dependencies for code execution.

chore: Add debug regex script for testing XML parsing

- Introduced `debug_regex.py` to validate regex patterns against XML tool calls.

chore: Add HTML template for displaying thinking stream events

- Created `test_thinking_stream.html` for visualizing tool calling events in a user-friendly format.

test: Add tests for OllamaAdapter XML parsing

- Developed `test_ollama_parser.py` to validate XML parsing with various test cases, including malformed XML.
This commit is contained in:
serversdwn
2025-12-26 03:49:20 -05:00
parent f1471cde84
commit 64429b19e6
37 changed files with 3238 additions and 23 deletions

View File

@@ -0,0 +1,191 @@
"""
Ollama adapter for tool calling using XML-structured prompts.
Since Ollama doesn't have native function calling, this adapter uses
XML-based prompts to instruct the model how to call tools.
"""
import json
import re
from typing import Dict, List, Optional
from .base import ToolAdapter
class OllamaAdapter(ToolAdapter):
"""Ollama adapter using XML-structured prompts for tool calling.
This adapter injects tool descriptions into the system prompt and
teaches the model to respond with XML when it wants to use a tool.
"""
SYSTEM_PROMPT = """You have access to the following tools:
{tool_descriptions}
To use a tool, respond with XML in this exact format:
<tool_call>
<name>tool_name</name>
<arguments>
<arg_name>value</arg_name>
</arguments>
<reason>why you're using this tool</reason>
</tool_call>
You can call multiple tools by including multiple <tool_call> blocks.
If you don't need to use any tools, respond normally without XML.
After tools are executed, you'll receive results and can continue the conversation."""
async def prepare_request(
self,
messages: List[Dict],
tools: List[Dict],
tool_choice: Optional[str] = None
) -> Dict:
"""Inject tool descriptions into system prompt.
Args:
messages: Conversation history
tools: Lyra tool definitions
tool_choice: Ignored for Ollama (no native support)
Returns:
dict: Request payload with modified messages
"""
# Format tool descriptions
tool_desc = "\n".join([
f"- {t['name']}: {t['description']}\n Parameters: {self._format_parameters(t['parameters'], t.get('required', []))}"
for t in tools
])
system_msg = self.SYSTEM_PROMPT.format(tool_descriptions=tool_desc)
# Check if first message is already a system message
modified_messages = messages.copy()
if modified_messages and modified_messages[0].get("role") == "system":
# Prepend tool instructions to existing system message
modified_messages[0]["content"] = system_msg + "\n\n" + modified_messages[0]["content"]
else:
# Add new system message at the beginning
modified_messages.insert(0, {"role": "system", "content": system_msg})
return {"messages": modified_messages}
def _format_parameters(self, parameters: Dict, required: List[str]) -> str:
"""Format parameters for tool description.
Args:
parameters: Parameter definitions
required: List of required parameter names
Returns:
str: Human-readable parameter description
"""
param_strs = []
for name, spec in parameters.items():
req_marker = "(required)" if name in required else "(optional)"
param_strs.append(f"{name} {req_marker}: {spec.get('description', '')}")
return ", ".join(param_strs)
async def parse_response(self, response) -> Dict:
"""Extract tool calls from XML in response.
Args:
response: String response from Ollama
Returns:
dict: Standardized Lyra format with content and tool_calls
"""
import logging
logger = logging.getLogger(__name__)
# Ollama returns a string
if isinstance(response, dict):
content = response.get("message", {}).get("content", "")
else:
content = str(response)
logger.info(f"🔍 OllamaAdapter.parse_response: content length={len(content)}, has <tool_call>={('<tool_call>' in content)}")
logger.debug(f"🔍 Content preview: {content[:500]}")
# Parse XML tool calls
tool_calls = []
if "<tool_call>" in content:
# Split content by <tool_call> to get each block
blocks = content.split('<tool_call>')
logger.info(f"🔍 Split into {len(blocks)} blocks")
# First block is content before any tool calls
clean_parts = [blocks[0]]
for idx, block in enumerate(blocks[1:]): # Skip first block (pre-tool content)
# Extract tool name
name_match = re.search(r'<name>(.*?)</name>', block)
if not name_match:
logger.warning(f"Block {idx} has no <name> tag, skipping")
continue
name = name_match.group(1).strip()
arguments = {}
# Extract arguments
args_match = re.search(r'<arguments>(.*?)</arguments>', block, re.DOTALL)
if args_match:
args_xml = args_match.group(1)
# Parse <key>value</key> pairs
arg_pairs = re.findall(r'<(\w+)>(.*?)</\1>', args_xml, re.DOTALL)
arguments = {k: v.strip() for k, v in arg_pairs}
tool_calls.append({
"id": f"call_{idx}",
"name": name,
"arguments": arguments
})
# For clean content, find what comes AFTER the tool call block
# Look for the last closing tag (</tool_call> or malformed </xxx>) and keep what's after
# Split by any closing tag at the END of the tool block
remaining = block
# Remove everything up to and including a standalone closing tag
# Pattern: find </something> that's not followed by more XML
end_match = re.search(r'</[a-z_]+>\s*(.*)$', remaining, re.DOTALL)
if end_match:
after_content = end_match.group(1).strip()
if after_content and not after_content.startswith('<'):
# Only keep if it's actual text content, not more XML
clean_parts.append(after_content)
clean_content = ''.join(clean_parts).strip()
else:
clean_content = content
return {
"content": clean_content,
"tool_calls": tool_calls if tool_calls else None
}
def format_tool_result(
self,
tool_call_id: str,
tool_name: str,
result: Dict
) -> Dict:
"""Format tool result as XML for next prompt.
Args:
tool_call_id: ID from the original tool call
tool_name: Name of the executed tool
result: Tool execution result
Returns:
dict: Message in user role with XML-formatted result
"""
# Format result as XML
result_xml = f"""<tool_result>
<tool>{tool_name}</tool>
<result>{json.dumps(result, ensure_ascii=False)}</result>
</tool_result>"""
return {
"role": "user",
"content": result_xml
}