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
@@ -3,6 +3,8 @@ import os
|
||||
import httpx
|
||||
import json
|
||||
import logging
|
||||
from typing import Optional, List, Dict
|
||||
from autonomy.tools.adapters import OpenAIAdapter, OllamaAdapter, LlamaCppAdapter
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
@@ -39,6 +41,16 @@ DEFAULT_BACKEND = "PRIMARY"
|
||||
# Reusable async HTTP client
|
||||
http_client = httpx.AsyncClient(timeout=120.0)
|
||||
|
||||
# Tool adapters for each backend
|
||||
TOOL_ADAPTERS = {
|
||||
"OPENAI": OpenAIAdapter(),
|
||||
"OLLAMA": OllamaAdapter(),
|
||||
"MI50": LlamaCppAdapter(), # MI50 uses llama.cpp
|
||||
"PRIMARY": None, # Determined at runtime
|
||||
"SECONDARY": None, # Determined at runtime
|
||||
"FALLBACK": None, # Determined at runtime
|
||||
}
|
||||
|
||||
|
||||
# ------------------------------------------------------------
|
||||
# Public call
|
||||
@@ -49,9 +61,12 @@ async def call_llm(
|
||||
backend: str | None = None,
|
||||
temperature: float = 0.7,
|
||||
max_tokens: int = 512,
|
||||
tools: Optional[List[Dict]] = None,
|
||||
tool_choice: Optional[str] = None,
|
||||
return_adapter_response: bool = False,
|
||||
):
|
||||
"""
|
||||
Call an LLM backend.
|
||||
Call an LLM backend with optional tool calling support.
|
||||
|
||||
Args:
|
||||
prompt: String prompt (for completion-style APIs like mi50)
|
||||
@@ -59,6 +74,13 @@ async def call_llm(
|
||||
backend: Which backend to use (PRIMARY, SECONDARY, OPENAI, etc.)
|
||||
temperature: Sampling temperature
|
||||
max_tokens: Maximum tokens to generate
|
||||
tools: List of Lyra tool definitions (provider-agnostic)
|
||||
tool_choice: How to use tools ("auto", "required", "none")
|
||||
return_adapter_response: If True, return dict with content and tool_calls
|
||||
|
||||
Returns:
|
||||
str (default) or dict (if return_adapter_response=True):
|
||||
{"content": str, "tool_calls": [...] or None}
|
||||
"""
|
||||
backend = (backend or DEFAULT_BACKEND).upper()
|
||||
|
||||
@@ -73,10 +95,54 @@ async def call_llm(
|
||||
if not url or not model:
|
||||
raise RuntimeError(f"Backend '{backend}' missing url/model in env")
|
||||
|
||||
# If tools are requested, use adapter to prepare request
|
||||
if tools:
|
||||
# Get adapter for this backend
|
||||
adapter = TOOL_ADAPTERS.get(backend)
|
||||
|
||||
# For PRIMARY/SECONDARY/FALLBACK, determine adapter based on provider
|
||||
if adapter is None and backend in ["PRIMARY", "SECONDARY", "FALLBACK"]:
|
||||
if provider == "openai":
|
||||
adapter = TOOL_ADAPTERS["OPENAI"]
|
||||
elif provider == "ollama":
|
||||
adapter = TOOL_ADAPTERS["OLLAMA"]
|
||||
elif provider == "mi50":
|
||||
adapter = TOOL_ADAPTERS["MI50"]
|
||||
|
||||
if adapter:
|
||||
# Use messages array if provided, otherwise convert prompt to messages
|
||||
if not messages:
|
||||
messages = [{"role": "user", "content": prompt}]
|
||||
|
||||
# Prepare request through adapter
|
||||
adapted_request = await adapter.prepare_request(messages, tools, tool_choice)
|
||||
messages = adapted_request["messages"]
|
||||
|
||||
# Extract tools in provider format if present
|
||||
provider_tools = adapted_request.get("tools")
|
||||
provider_tool_choice = adapted_request.get("tool_choice")
|
||||
else:
|
||||
logger.warning(f"No adapter available for backend {backend}, ignoring tools")
|
||||
provider_tools = None
|
||||
provider_tool_choice = None
|
||||
else:
|
||||
provider_tools = None
|
||||
provider_tool_choice = None
|
||||
|
||||
# -------------------------------
|
||||
# Provider: MI50 (llama.cpp server)
|
||||
# -------------------------------
|
||||
if provider == "mi50":
|
||||
# If tools requested, convert messages to prompt with tool instructions
|
||||
if messages and tools:
|
||||
# Combine messages into a prompt
|
||||
prompt_parts = []
|
||||
for msg in messages:
|
||||
role = msg.get("role", "user")
|
||||
content = msg.get("content", "")
|
||||
prompt_parts.append(f"{role.capitalize()}: {content}")
|
||||
prompt = "\n".join(prompt_parts) + "\nAssistant:"
|
||||
|
||||
payload = {
|
||||
"prompt": prompt,
|
||||
"n_predict": max_tokens,
|
||||
@@ -87,7 +153,15 @@ async def call_llm(
|
||||
r = await http_client.post(f"{url}/completion", json=payload)
|
||||
r.raise_for_status()
|
||||
data = r.json()
|
||||
return data.get("content", "")
|
||||
response_content = data.get("content", "")
|
||||
|
||||
# If caller wants adapter response with tool calls, parse and return
|
||||
if return_adapter_response and tools:
|
||||
adapter = TOOL_ADAPTERS.get(backend) or TOOL_ADAPTERS["MI50"]
|
||||
return await adapter.parse_response(response_content)
|
||||
else:
|
||||
return response_content
|
||||
|
||||
except httpx.HTTPError as e:
|
||||
logger.error(f"HTTP error calling mi50: {type(e).__name__}: {str(e)}")
|
||||
raise RuntimeError(f"LLM API error (mi50): {type(e).__name__}: {str(e)}")
|
||||
@@ -101,7 +175,9 @@ async def call_llm(
|
||||
# -------------------------------
|
||||
# Provider: OLLAMA (your 3090)
|
||||
# -------------------------------
|
||||
logger.info(f"🔍 LLM Router: provider={provider}, checking if ollama...")
|
||||
if provider == "ollama":
|
||||
logger.info(f"🔍 LLM Router: Matched ollama provider, tools={bool(tools)}, return_adapter_response={return_adapter_response}")
|
||||
# Use messages array if provided, otherwise convert prompt to single user message
|
||||
if messages:
|
||||
chat_messages = messages
|
||||
@@ -121,7 +197,19 @@ async def call_llm(
|
||||
r = await http_client.post(f"{url}/api/chat", json=payload)
|
||||
r.raise_for_status()
|
||||
data = r.json()
|
||||
return data["message"]["content"]
|
||||
response_content = data["message"]["content"]
|
||||
|
||||
# If caller wants adapter response with tool calls, parse and return
|
||||
if return_adapter_response and tools:
|
||||
logger.info(f"🔍 Ollama: return_adapter_response=True, calling adapter.parse_response")
|
||||
adapter = TOOL_ADAPTERS.get(backend) or TOOL_ADAPTERS["OLLAMA"]
|
||||
logger.info(f"🔍 Ollama: Using adapter {adapter.__class__.__name__}")
|
||||
result = await adapter.parse_response(response_content)
|
||||
logger.info(f"🔍 Ollama: Adapter returned {result}")
|
||||
return result
|
||||
else:
|
||||
return response_content
|
||||
|
||||
except httpx.HTTPError as e:
|
||||
logger.error(f"HTTP error calling ollama: {type(e).__name__}: {str(e)}")
|
||||
raise RuntimeError(f"LLM API error (ollama): {type(e).__name__}: {str(e)}")
|
||||
@@ -154,11 +242,49 @@ async def call_llm(
|
||||
"temperature": temperature,
|
||||
"max_tokens": max_tokens,
|
||||
}
|
||||
|
||||
# Add tools if available (OpenAI native function calling)
|
||||
if provider_tools:
|
||||
payload["tools"] = provider_tools
|
||||
if provider_tool_choice:
|
||||
payload["tool_choice"] = provider_tool_choice
|
||||
|
||||
try:
|
||||
r = await http_client.post(f"{url}/chat/completions", json=payload, headers=headers)
|
||||
r.raise_for_status()
|
||||
data = r.json()
|
||||
return data["choices"][0]["message"]["content"]
|
||||
|
||||
# If caller wants adapter response with tool calls, parse and return
|
||||
if return_adapter_response and tools:
|
||||
# Create mock response object for adapter
|
||||
class MockChoice:
|
||||
def __init__(self, message_data):
|
||||
self.message = type('obj', (object,), {})()
|
||||
self.message.content = message_data.get("content")
|
||||
# Convert tool_calls dicts to objects
|
||||
raw_tool_calls = message_data.get("tool_calls")
|
||||
if raw_tool_calls:
|
||||
self.message.tool_calls = []
|
||||
for tc in raw_tool_calls:
|
||||
tool_call_obj = type('obj', (object,), {})()
|
||||
tool_call_obj.id = tc.get("id")
|
||||
tool_call_obj.function = type('obj', (object,), {})()
|
||||
tool_call_obj.function.name = tc.get("function", {}).get("name")
|
||||
tool_call_obj.function.arguments = tc.get("function", {}).get("arguments")
|
||||
self.message.tool_calls.append(tool_call_obj)
|
||||
else:
|
||||
self.message.tool_calls = None
|
||||
|
||||
class MockResponse:
|
||||
def __init__(self, data):
|
||||
self.choices = [MockChoice(data["choices"][0]["message"])]
|
||||
|
||||
mock_resp = MockResponse(data)
|
||||
adapter = TOOL_ADAPTERS.get(backend) or TOOL_ADAPTERS["OPENAI"]
|
||||
return await adapter.parse_response(mock_resp)
|
||||
else:
|
||||
return data["choices"][0]["message"]["content"]
|
||||
|
||||
except httpx.HTTPError as e:
|
||||
logger.error(f"HTTP error calling openai: {type(e).__name__}: {str(e)}")
|
||||
raise RuntimeError(f"LLM API error (openai): {type(e).__name__}: {str(e)}")
|
||||
|
||||
Reference in New Issue
Block a user