feat!: migrate to OpenAI native tool calls and fix circular deps (#fuck-gemini)
- Fix circular dependencies in agent/tools - Migrate from custom JSON to OpenAI tool calls format - Add async streaming (step_stream, complete_stream) - Simplify prompt system and remove token counting - Add 5 new API endpoints (/health, /v1/models, /api/memory/*) - Add 3 new tools (get_torrent_by_index, add_torrent_by_index, set_language) - Fix all 500 tests and add coverage config (80% threshold) - Add comprehensive docs (README, pytest guide) BREAKING: LLM interface changed, memory injection via get_memory()
This commit is contained in:
+138
-56
@@ -1,15 +1,27 @@
|
||||
# agent/prompts.py
|
||||
from typing import Dict, Any
|
||||
"""Prompt builder for the agent system."""
|
||||
|
||||
import json
|
||||
|
||||
from .registry import Tool
|
||||
from infrastructure.persistence import get_memory
|
||||
|
||||
from .parameters import format_parameters_for_prompt, get_missing_required_parameters
|
||||
from .registry import Tool
|
||||
|
||||
|
||||
class PromptBuilder:
|
||||
"""Handles construction of system prompts for the agent."""
|
||||
"""Builds system prompts for the agent with memory context.
|
||||
|
||||
def __init__(self, tools: Dict[str, Tool]):
|
||||
Attributes:
|
||||
tools: Dictionary of available tools.
|
||||
"""
|
||||
|
||||
def __init__(self, tools: dict[str, Tool]):
|
||||
"""
|
||||
Initialize the prompt builder.
|
||||
|
||||
Args:
|
||||
tools: Dictionary mapping tool names to Tool instances.
|
||||
"""
|
||||
self.tools = tools
|
||||
|
||||
def _format_tools_description(self) -> str:
|
||||
@@ -20,69 +32,139 @@ class PromptBuilder:
|
||||
for tool in self.tools.values()
|
||||
)
|
||||
|
||||
def _build_context(self, memory_data: dict) -> Dict[str, Any]:
|
||||
"""Build the context object with current state from memory."""
|
||||
return memory_data
|
||||
def _format_episodic_context(self) -> str:
|
||||
"""Format episodic memory context for the prompt."""
|
||||
memory = get_memory()
|
||||
lines = []
|
||||
|
||||
def build_system_prompt(self, memory_data: dict) -> str:
|
||||
# Last search results
|
||||
if memory.episodic.last_search_results:
|
||||
search = memory.episodic.last_search_results
|
||||
lines.append(f"LAST SEARCH: '{search.get('query')}'")
|
||||
results = search.get("results", [])
|
||||
if results:
|
||||
lines.append(f" {len(results)} results available:")
|
||||
for r in results[:5]:
|
||||
name = r.get("name", r.get("title", "Unknown"))
|
||||
lines.append(f" {r.get('index')}. {name}")
|
||||
if len(results) > 5:
|
||||
lines.append(f" ... and {len(results) - 5} more")
|
||||
|
||||
# Pending question
|
||||
if memory.episodic.pending_question:
|
||||
q = memory.episodic.pending_question
|
||||
lines.append(f"\nPENDING QUESTION: {q.get('question')}")
|
||||
for opt in q.get("options", []):
|
||||
lines.append(f" {opt.get('index')}. {opt.get('label')}")
|
||||
|
||||
# Active downloads
|
||||
if memory.episodic.active_downloads:
|
||||
lines.append(f"\nACTIVE DOWNLOADS: {len(memory.episodic.active_downloads)}")
|
||||
for dl in memory.episodic.active_downloads[:3]:
|
||||
lines.append(f" - {dl.get('name')}: {dl.get('progress', 0)}%")
|
||||
|
||||
# Recent errors
|
||||
if memory.episodic.recent_errors:
|
||||
last_error = memory.episodic.recent_errors[-1]
|
||||
lines.append(
|
||||
f"\nLAST ERROR: {last_error.get('error')} "
|
||||
f"(action: {last_error.get('action')})"
|
||||
)
|
||||
|
||||
# Unread events
|
||||
unread = [e for e in memory.episodic.background_events if not e.get("read")]
|
||||
if unread:
|
||||
lines.append(f"\nUNREAD EVENTS: {len(unread)}")
|
||||
for e in unread[:3]:
|
||||
lines.append(f" - {e.get('type')}: {e.get('data', {})}")
|
||||
|
||||
return "\n".join(lines) if lines else ""
|
||||
|
||||
def _format_stm_context(self) -> str:
|
||||
"""Format short-term memory context for the prompt."""
|
||||
memory = get_memory()
|
||||
lines = []
|
||||
|
||||
# Current workflow
|
||||
if memory.stm.current_workflow:
|
||||
wf = memory.stm.current_workflow
|
||||
lines.append(f"CURRENT WORKFLOW: {wf.get('type')}")
|
||||
lines.append(f" Target: {wf.get('target', {}).get('title', 'Unknown')}")
|
||||
lines.append(f" Stage: {wf.get('stage')}")
|
||||
|
||||
# Current topic
|
||||
if memory.stm.current_topic:
|
||||
lines.append(f"CURRENT TOPIC: {memory.stm.current_topic}")
|
||||
|
||||
# Extracted entities
|
||||
if memory.stm.extracted_entities:
|
||||
entities_json = json.dumps(
|
||||
memory.stm.extracted_entities, ensure_ascii=False
|
||||
)
|
||||
lines.append(f"EXTRACTED ENTITIES: {entities_json}")
|
||||
|
||||
return "\n".join(lines) if lines else ""
|
||||
|
||||
def build_system_prompt(self) -> str:
|
||||
"""
|
||||
Build the system prompt with context provided as JSON.
|
||||
|
||||
Args:
|
||||
memory_data: The full memory data dictionary
|
||||
Build the system prompt with context from memory.
|
||||
|
||||
Returns:
|
||||
The complete system prompt string
|
||||
The complete system prompt string.
|
||||
"""
|
||||
context = self._build_context(memory_data)
|
||||
memory = get_memory()
|
||||
tools_desc = self._format_tools_description()
|
||||
params_desc = format_parameters_for_prompt()
|
||||
|
||||
# Check for missing required parameters
|
||||
missing_params = get_missing_required_parameters(memory_data)
|
||||
missing_params = get_missing_required_parameters({"config": memory.ltm.config})
|
||||
missing_info = ""
|
||||
if missing_params:
|
||||
missing_info = "\n\n⚠️ MISSING REQUIRED PARAMETERS:\n"
|
||||
missing_info = "\n\nMISSING REQUIRED PARAMETERS:\n"
|
||||
for param in missing_params:
|
||||
missing_info += f"- {param.key}: {param.description}\n"
|
||||
missing_info += f" Why needed: {param.why_needed}\n"
|
||||
|
||||
return (
|
||||
"You are an AI agent helping a user manage their local media library.\n\n"
|
||||
f"{params_desc}\n\n"
|
||||
"CURRENT CONTEXT (JSON):\n"
|
||||
f"{json.dumps(context, indent=2, ensure_ascii=False)}\n"
|
||||
f"{missing_info}\n"
|
||||
"IMPORTANT RULES:\n"
|
||||
"1. Check the REQUIRED PARAMETERS section above to understand what information you need.\n"
|
||||
"2. If any required parameter is missing (shown in MISSING REQUIRED PARAMETERS), "
|
||||
"you MUST ask the user for it and explain WHY you need it based on the parameter description.\n"
|
||||
"3. To use a tool, respond STRICTLY with this JSON format:\n"
|
||||
' { "thought": "explanation", "action": { "name": "tool_name", "args": { "arg": "value" } } }\n'
|
||||
" - No text before or after the JSON\n"
|
||||
" - All args must be complete and non-null\n"
|
||||
"4. You can use MULTIPLE TOOLS IN SEQUENCE:\n"
|
||||
" - After executing a tool, you will receive its result\n"
|
||||
" - You can then decide to use another tool based on the result\n"
|
||||
" - Or provide a final text response to the user\n"
|
||||
" - Continue using tools until you have all the information needed\n"
|
||||
"5. If you respond with text (not using a tool), respond normally in French.\n"
|
||||
"6. When you have all the information needed, provide a final response in NATURAL TEXT (not JSON).\n"
|
||||
"7. Extract the relevant information from the user's request and pass it as tool arguments.\n"
|
||||
"\n"
|
||||
"EXAMPLES:\n"
|
||||
" To set the download folder:\n"
|
||||
' { "thought": "User provided download path", "action": { "name": "set_path", "args": { "path_type": "download_folder", "path_value": "/home/user/downloads" } } }\n'
|
||||
"\n"
|
||||
" To set the TV show folder:\n"
|
||||
' { "thought": "User provided TV show path", "action": { "name": "set_path", "args": { "path_type": "tvshow_folder", "path_value": "/home/user/media/tvshows" } } }\n'
|
||||
"\n"
|
||||
" To list the download folder:\n"
|
||||
' { "thought": "User wants to see downloads", "action": { "name": "list_folder", "args": { "folder_type": "download", "path": "." } } }\n'
|
||||
"\n"
|
||||
" To list a subfolder in TV shows:\n"
|
||||
' { "thought": "User wants to see a specific show", "action": { "name": "list_folder", "args": { "folder_type": "tvshow", "path": "Game.of.Thrones" } } }\n'
|
||||
"\n"
|
||||
"AVAILABLE TOOLS:\n"
|
||||
f"{tools_desc}\n"
|
||||
)
|
||||
# Build context sections
|
||||
episodic_context = self._format_episodic_context()
|
||||
stm_context = self._format_stm_context()
|
||||
|
||||
config_json = json.dumps(memory.ltm.config, indent=2, ensure_ascii=False)
|
||||
|
||||
return f"""You are an AI agent helping a user manage their local media library.
|
||||
|
||||
{params_desc}
|
||||
|
||||
CURRENT CONFIGURATION:
|
||||
{config_json}
|
||||
{missing_info}
|
||||
|
||||
{f"SESSION CONTEXT:{chr(10)}{stm_context}" if stm_context else ""}
|
||||
|
||||
{f"CURRENT STATE:{chr(10)}{episodic_context}" if episodic_context else ""}
|
||||
|
||||
IMPORTANT RULES:
|
||||
1. When the user refers to a number (e.g., "the 3rd one", "download number 2"), \
|
||||
use `add_torrent_by_index` or `get_torrent_by_index` with that number.
|
||||
2. If a torrent search was performed, results are numbered. \
|
||||
The user can reference them by number.
|
||||
3. To use a tool, respond STRICTLY with this JSON format:
|
||||
{{ "thought": "explanation", "action": {{ "name": "tool_name", "args": {{ }} }} }}
|
||||
- No text before or after the JSON
|
||||
4. You can use MULTIPLE TOOLS IN SEQUENCE.
|
||||
5. When you have all the information needed, respond in NATURAL TEXT (not JSON).
|
||||
6. If a required parameter is missing, ask the user for it.
|
||||
7. Respond in the same language as the user.
|
||||
|
||||
EXAMPLES:
|
||||
- After a torrent search, if the user says "download the 3rd one":
|
||||
{{ "thought": "User wants torrent #3", "action": {{ "name": "add_torrent_by_index", \
|
||||
"args": {{ "index": 3 }} }} }}
|
||||
|
||||
- To search for torrents:
|
||||
{{ "thought": "Searching torrents", "action": {{ "name": "find_torrents", \
|
||||
"args": {{ "media_title": "Inception 1080p" }} }} }}
|
||||
|
||||
AVAILABLE TOOLS:
|
||||
{tools_desc}
|
||||
"""
|
||||
|
||||
Reference in New Issue
Block a user