Unfucked gemini's mess
This commit is contained in:
+152
-237
@@ -1,8 +1,7 @@
|
||||
"""Main agent for media library management."""
|
||||
|
||||
import json
|
||||
import logging
|
||||
from typing import Any, Protocol
|
||||
from typing import Any, Dict, List, Optional
|
||||
|
||||
from infrastructure.persistence import get_memory
|
||||
|
||||
@@ -13,266 +12,182 @@ from .registry import Tool, make_tools
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class LLMClient(Protocol):
|
||||
"""Protocol defining the LLM client interface."""
|
||||
|
||||
def complete(self, messages: list[dict[str, Any]]) -> str:
|
||||
"""Send messages to the LLM and get a response."""
|
||||
...
|
||||
|
||||
|
||||
class Agent:
|
||||
"""
|
||||
AI agent for media library management.
|
||||
|
||||
Orchestrates interactions between the LLM, memory, and tools
|
||||
to respond to user requests.
|
||||
|
||||
Attributes:
|
||||
llm: LLM client (DeepSeek or Ollama).
|
||||
tools: Available tools for the agent.
|
||||
prompt_builder: Builds system prompts with context.
|
||||
max_tool_iterations: Maximum tool calls per request.
|
||||
|
||||
Uses OpenAI-compatible tool calling API.
|
||||
"""
|
||||
|
||||
def __init__(self, llm: LLMClient, max_tool_iterations: int = 5):
|
||||
def __init__(self, llm, max_tool_iterations: int = 5):
|
||||
"""
|
||||
Initialize the agent.
|
||||
|
||||
|
||||
Args:
|
||||
llm: LLM client compatible with the LLMClient protocol.
|
||||
max_tool_iterations: Maximum tool iterations (default: 5).
|
||||
llm: LLM client with complete() method
|
||||
max_tool_iterations: Maximum number of tool execution iterations
|
||||
"""
|
||||
self.llm = llm
|
||||
self.tools: dict[str, Tool] = make_tools()
|
||||
self.tools: Dict[str, Tool] = make_tools()
|
||||
self.prompt_builder = PromptBuilder(self.tools)
|
||||
self.max_tool_iterations = max_tool_iterations
|
||||
|
||||
def _parse_intent(self, text: str) -> dict[str, Any] | None:
|
||||
"""
|
||||
Parse an LLM response to detect a tool call.
|
||||
|
||||
Args:
|
||||
text: LLM response text.
|
||||
|
||||
Returns:
|
||||
Dict with intent if a tool call is detected, None otherwise.
|
||||
"""
|
||||
text = text.strip()
|
||||
|
||||
# Try direct JSON parse
|
||||
if text.startswith("{") and text.endswith("}"):
|
||||
try:
|
||||
data = json.loads(text)
|
||||
if self._is_valid_intent(data):
|
||||
return data
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
|
||||
# Try to extract JSON from text
|
||||
try:
|
||||
start = text.find("{")
|
||||
end = text.rfind("}") + 1
|
||||
if start != -1 and end > start:
|
||||
json_str = text[start:end]
|
||||
data = json.loads(json_str)
|
||||
if self._is_valid_intent(data):
|
||||
return data
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
|
||||
return None
|
||||
|
||||
def _is_valid_intent(self, data: Any) -> bool:
|
||||
"""Check if parsed data is a valid tool intent."""
|
||||
if not isinstance(data, dict) or "action" not in data:
|
||||
return False
|
||||
action = data.get("action")
|
||||
return isinstance(action, dict) and isinstance(action.get("name"), str)
|
||||
|
||||
def _execute_action(self, intent: dict[str, Any]) -> dict[str, Any]:
|
||||
"""
|
||||
Execute a tool action requested by the LLM.
|
||||
|
||||
Args:
|
||||
intent: Dict containing the action to execute.
|
||||
|
||||
Returns:
|
||||
Tool execution result.
|
||||
"""
|
||||
action = intent["action"]
|
||||
name: str = action["name"]
|
||||
args: dict[str, Any] = action.get("args", {}) or {}
|
||||
|
||||
tool = self.tools.get(name)
|
||||
if not tool:
|
||||
logger.warning(f"Unknown tool requested: {name}")
|
||||
return {
|
||||
"error": "unknown_tool",
|
||||
"tool": name,
|
||||
"available_tools": list(self.tools.keys()),
|
||||
}
|
||||
|
||||
try:
|
||||
result = tool.func(**args)
|
||||
|
||||
# Track errors in episodic memory
|
||||
if result.get("status") == "error" or result.get("error"):
|
||||
memory = get_memory()
|
||||
memory.episodic.add_error(
|
||||
action=name,
|
||||
error=result.get("error", result.get("message", "Unknown error")),
|
||||
context={"args": args, "result": result},
|
||||
)
|
||||
|
||||
return result
|
||||
|
||||
except TypeError as e:
|
||||
error_msg = f"Bad arguments for {name}: {e}"
|
||||
logger.error(error_msg)
|
||||
memory = get_memory()
|
||||
memory.episodic.add_error(
|
||||
action=name, error=error_msg, context={"args": args}
|
||||
)
|
||||
return {"error": "bad_args", "message": str(e)}
|
||||
|
||||
except Exception as e:
|
||||
error_msg = f"Error executing {name}: {e}"
|
||||
logger.error(error_msg, exc_info=True)
|
||||
memory = get_memory()
|
||||
memory.episodic.add_error(action=name, error=str(e), context={"args": args})
|
||||
return {"error": "execution_error", "message": str(e)}
|
||||
|
||||
def _check_unread_events(self) -> str:
|
||||
"""
|
||||
Check for unread background events and format them.
|
||||
|
||||
Returns:
|
||||
Formatted string of events, or empty string if none.
|
||||
"""
|
||||
memory = get_memory()
|
||||
events = memory.episodic.get_unread_events()
|
||||
|
||||
if not events:
|
||||
return ""
|
||||
|
||||
lines = ["Recent events:"]
|
||||
for event in events:
|
||||
event_type = event.get("type", "unknown")
|
||||
data = event.get("data", {})
|
||||
|
||||
if event_type == "download_complete":
|
||||
lines.append(f" - Download completed: {data.get('name')}")
|
||||
elif event_type == "new_files_detected":
|
||||
lines.append(f" - {data.get('count')} new files detected")
|
||||
else:
|
||||
lines.append(f" - {event_type}: {data}")
|
||||
|
||||
return "\n".join(lines)
|
||||
|
||||
def step(self, user_input: str) -> str:
|
||||
"""
|
||||
Execute one agent step with iterative tool execution.
|
||||
|
||||
Process:
|
||||
1. Check for unread events
|
||||
2. Build system prompt with memory context
|
||||
3. Query the LLM
|
||||
4. If tool call detected, execute and loop
|
||||
5. Return final text response
|
||||
|
||||
Execute one agent step with the user input.
|
||||
|
||||
This method:
|
||||
1. Adds user message to memory
|
||||
2. Builds prompt with history and context
|
||||
3. Calls LLM, executing tools as needed
|
||||
4. Returns final response
|
||||
|
||||
Args:
|
||||
user_input: User message.
|
||||
|
||||
user_input: User's message
|
||||
|
||||
Returns:
|
||||
Final response in natural text.
|
||||
Agent's final response
|
||||
"""
|
||||
logger.info("Starting agent step")
|
||||
logger.debug(f"User input: {user_input}")
|
||||
|
||||
memory = get_memory()
|
||||
|
||||
# Check for background events
|
||||
events_notification = self._check_unread_events()
|
||||
if events_notification:
|
||||
logger.info("Found unread background events")
|
||||
|
||||
# Build system prompt
|
||||
|
||||
# Add user message to history
|
||||
memory.stm.add_message("user", user_input)
|
||||
memory.save()
|
||||
|
||||
# Build initial messages
|
||||
system_prompt = self.prompt_builder.build_system_prompt()
|
||||
|
||||
# Initialize conversation
|
||||
messages: list[dict[str, Any]] = [
|
||||
{"role": "system", "content": system_prompt},
|
||||
messages: List[Dict[str, Any]] = [
|
||||
{"role": "system", "content": system_prompt}
|
||||
]
|
||||
|
||||
|
||||
# Add conversation history
|
||||
history = memory.stm.get_recent_history(settings.max_history_messages)
|
||||
if history:
|
||||
for msg in history:
|
||||
messages.append({"role": msg["role"], "content": msg["content"]})
|
||||
logger.debug(f"Added {len(history)} messages from history")
|
||||
|
||||
# Add events notification
|
||||
if events_notification:
|
||||
messages.append(
|
||||
{"role": "system", "content": f"[NOTIFICATION]\n{events_notification}"}
|
||||
)
|
||||
|
||||
# Add user input
|
||||
messages.append({"role": "user", "content": user_input})
|
||||
|
||||
messages.extend(history)
|
||||
|
||||
# Add unread events if any
|
||||
unread_events = memory.episodic.get_unread_events()
|
||||
if unread_events:
|
||||
events_text = "\n".join([
|
||||
f"- {e['type']}: {e['data']}"
|
||||
for e in unread_events
|
||||
])
|
||||
messages.append({
|
||||
"role": "system",
|
||||
"content": f"Background events:\n{events_text}"
|
||||
})
|
||||
|
||||
# Get tools specification for OpenAI format
|
||||
tools_spec = self.prompt_builder.build_tools_spec()
|
||||
|
||||
# Tool execution loop
|
||||
iteration = 0
|
||||
while iteration < self.max_tool_iterations:
|
||||
logger.debug(f"Iteration {iteration + 1}/{self.max_tool_iterations}")
|
||||
|
||||
llm_response = self.llm.complete(messages)
|
||||
logger.debug(f"LLM response: {llm_response[:200]}...")
|
||||
|
||||
intent = self._parse_intent(llm_response)
|
||||
|
||||
if not intent:
|
||||
# Final text response
|
||||
logger.info("No tool intent, returning response")
|
||||
memory.stm.add_message("user", user_input)
|
||||
memory.stm.add_message("assistant", llm_response)
|
||||
for iteration in range(self.max_tool_iterations):
|
||||
# Call LLM with tools
|
||||
llm_result = self.llm.complete(messages, tools=tools_spec)
|
||||
|
||||
# Handle both tuple (response, usage) and dict response
|
||||
if isinstance(llm_result, tuple):
|
||||
response_message, usage = llm_result
|
||||
else:
|
||||
response_message = llm_result
|
||||
|
||||
# Check if there are tool calls
|
||||
tool_calls = response_message.get("tool_calls")
|
||||
|
||||
if not tool_calls:
|
||||
# No tool calls, this is the final response
|
||||
final_content = response_message.get("content", "")
|
||||
memory.stm.add_message("assistant", final_content)
|
||||
memory.save()
|
||||
return llm_response
|
||||
|
||||
# Execute tool
|
||||
tool_name = intent.get("action", {}).get("name", "unknown")
|
||||
logger.info(f"Executing tool: {tool_name}")
|
||||
tool_result = self._execute_action(intent)
|
||||
logger.debug(f"Tool result: {tool_result}")
|
||||
|
||||
# Add to conversation
|
||||
messages.append(
|
||||
{"role": "assistant", "content": json.dumps(intent, ensure_ascii=False)}
|
||||
)
|
||||
messages.append(
|
||||
{
|
||||
"role": "user",
|
||||
"content": json.dumps(
|
||||
{"tool_result": tool_result}, ensure_ascii=False
|
||||
),
|
||||
}
|
||||
)
|
||||
|
||||
iteration += 1
|
||||
|
||||
# Max iterations reached
|
||||
logger.warning(f"Max iterations ({self.max_tool_iterations}) reached")
|
||||
messages.append(
|
||||
{
|
||||
"role": "user",
|
||||
"content": "Please provide a final response based on the results.",
|
||||
}
|
||||
)
|
||||
|
||||
final_response = self.llm.complete(messages)
|
||||
|
||||
memory.stm.add_message("user", user_input)
|
||||
return final_content
|
||||
|
||||
# Add assistant message with tool calls to conversation
|
||||
messages.append(response_message)
|
||||
|
||||
# Execute each tool call
|
||||
for tool_call in tool_calls:
|
||||
tool_result = self._execute_tool_call(tool_call)
|
||||
|
||||
# Add tool result to messages
|
||||
messages.append({
|
||||
"tool_call_id": tool_call.get("id"),
|
||||
"role": "tool",
|
||||
"name": tool_call.get("function", {}).get("name"),
|
||||
"content": json.dumps(tool_result, ensure_ascii=False),
|
||||
})
|
||||
|
||||
# Max iterations reached, force final response
|
||||
messages.append({
|
||||
"role": "system",
|
||||
"content": "Please provide a final response to the user without using any more tools."
|
||||
})
|
||||
|
||||
llm_result = self.llm.complete(messages)
|
||||
if isinstance(llm_result, tuple):
|
||||
final_message, usage = llm_result
|
||||
else:
|
||||
final_message = llm_result
|
||||
|
||||
final_response = final_message.get("content", "I've completed the requested actions.")
|
||||
memory.stm.add_message("assistant", final_response)
|
||||
memory.save()
|
||||
|
||||
return final_response
|
||||
|
||||
def _execute_tool_call(self, tool_call: Dict[str, Any]) -> Dict[str, Any]:
|
||||
"""
|
||||
Execute a single tool call.
|
||||
|
||||
Args:
|
||||
tool_call: OpenAI-format tool call dict
|
||||
|
||||
Returns:
|
||||
Result dictionary
|
||||
"""
|
||||
function = tool_call.get("function", {})
|
||||
tool_name = function.get("name", "")
|
||||
|
||||
try:
|
||||
args_str = function.get("arguments", "{}")
|
||||
args = json.loads(args_str)
|
||||
except json.JSONDecodeError as e:
|
||||
logger.error(f"Failed to parse tool arguments: {e}")
|
||||
return {
|
||||
"error": "bad_args",
|
||||
"message": f"Invalid JSON arguments: {e}"
|
||||
}
|
||||
|
||||
# Validate tool exists
|
||||
if tool_name not in self.tools:
|
||||
available = list(self.tools.keys())
|
||||
return {
|
||||
"error": "unknown_tool",
|
||||
"message": f"Tool '{tool_name}' not found",
|
||||
"available_tools": available
|
||||
}
|
||||
|
||||
tool = self.tools[tool_name]
|
||||
|
||||
# Execute tool
|
||||
try:
|
||||
result = tool.func(**args)
|
||||
return result
|
||||
except KeyboardInterrupt:
|
||||
# Don't catch KeyboardInterrupt - let it propagate
|
||||
raise
|
||||
except TypeError as e:
|
||||
# Bad arguments
|
||||
memory = get_memory()
|
||||
memory.episodic.add_error(tool_name, f"bad_args: {e}")
|
||||
return {
|
||||
"error": "bad_args",
|
||||
"message": str(e),
|
||||
"tool": tool_name
|
||||
}
|
||||
except Exception as e:
|
||||
# Other errors
|
||||
memory = get_memory()
|
||||
memory.episodic.add_error(tool_name, str(e))
|
||||
return {
|
||||
"error": "execution_failed",
|
||||
"message": str(e),
|
||||
"tool": tool_name
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user