Files
alfred/alfred/agent/prompt.py
T

334 lines
12 KiB
Python

"""Prompt builder for the agent system."""
import json
from typing import Any
from alfred.infrastructure.persistence import get_memory
from alfred.infrastructure.persistence.memory import MemoryRegistry
from .expressions import build_expressions_context
from .registry import Tool
from .workflows import WorkflowLoader
# Tools that are always available, regardless of workflow scope.
# Kept small on purpose — the noyau is what the agent uses to either
# answer trivially or pivot into a workflow.
CORE_TOOLS: tuple[str, ...] = (
"set_language",
"set_path_for_folder",
"list_folder",
"read_release_metadata",
"query_library",
"start_workflow",
"end_workflow",
)
class PromptBuilder:
"""Builds system prompts for the agent with memory context."""
def __init__(
self,
tools: dict[str, Tool],
workflow_loader: WorkflowLoader | None = None,
):
self.tools = tools
self.workflow_loader = workflow_loader or WorkflowLoader()
self._memory_registry = MemoryRegistry()
def _active_workflow(self, memory) -> dict | None:
"""Return the YAML definition of the active workflow, or None."""
current = memory.stm.workflow.current
if current is None:
return None
return self.workflow_loader.get(current.get("name"))
def visible_tool_names(self) -> list[str]:
"""
Return the names of the tools currently in scope.
- Idle (no workflow): core noyau only. The LLM enters a workflow
via start_workflow to access more tools.
- Workflow active: core noyau + the workflow's declared tools.
"""
memory = get_memory()
visible = set(CORE_TOOLS)
workflow = self._active_workflow(memory)
if workflow is not None:
for name in workflow.get("tools", []):
visible.add(name)
# Only return tools that actually exist in the registry.
return [name for name in self.tools if name in visible]
def _format_identity(self, memory) -> str:
"""Build Alfred's identity and personality section."""
username = memory.stm.get_entity("username")
expr = build_expressions_context(username)
nickname = expr["nickname"]
samples = expr["samples"]
# Format expressions as situational guidance for the LLM
expr_lines = []
situation_labels = {
"greeting": "Salutation",
"success": "Succès",
"working": "En cours",
"error": "Erreur",
"unclear": "Demande floue",
"warning": "Avertissement",
"not_found": "Introuvable",
}
for situation, label in situation_labels.items():
phrase = samples.get(situation, "")
if phrase:
expr_lines.append(f' {label}: "{phrase}"')
expressions_block = "\n".join(expr_lines)
return f"""Tu t'appelles Alfred. Tu es un assistant d'organisation de médiathèque — direct, opérationnel, légèrement impertinent.
PERSONNALITÉ:
- Tu parles franglais : français avec des touches d'anglais quand ça colle mieux
- Tu es sarcastique et n'as pas peur de te moquer si l'user fait une connerie
- Tu envoies chier poliment (mais clairement) quand la demande est trop floue
- Tu ne fais pas de blabla inutile. Si "ok c'est fait" suffit, c'est tout ce que tu dis
- Tu peux jurer (putain, merde, con, ...) — c'est naturel, pas du remplissage
- Jamais de "Great question!" ou de politesse creuse
USER COURANT: {nickname}
EXPRESSIONS À UTILISER (une par situation, naturellement intégrées dans ta réponse) :
{expressions_block}"""
def build_tools_spec(self) -> list[dict[str, Any]]:
"""Build the tool specification for the LLM API (scope-filtered)."""
visible = set(self.visible_tool_names())
tool_specs = []
for tool in self.tools.values():
if tool.name not in visible:
continue
spec = {
"type": "function",
"function": {
"name": tool.name,
"description": tool.description,
"parameters": tool.parameters,
},
}
tool_specs.append(spec)
return tool_specs
def _format_tools_description(self) -> str:
"""Format the currently-visible tools with description + params."""
visible = set(self.visible_tool_names())
visible_tools = [t for t in self.tools.values() if t.name in visible]
if not visible_tools:
return ""
return "\n".join(
f"- {tool.name}: {tool.description}\n"
f" Parameters: {json.dumps(tool.parameters, ensure_ascii=False)}"
for tool in visible_tools
)
def _format_workflow_scope(self, memory) -> str:
"""Describe the current workflow scope so the LLM has a plan."""
workflow = self._active_workflow(memory)
if workflow is None:
available = self.workflow_loader.names()
if not available:
return ""
lines = ["WORKFLOW SCOPE: idle (broad catalog narrowed to core noyau)."]
lines.append(
" Call start_workflow(workflow_name, params) to enter a scope."
)
lines.append(" Available workflows:")
for name in available:
wf = self.workflow_loader.get(name) or {}
desc = (wf.get("description") or "").strip().splitlines()
summary = desc[0] if desc else ""
lines.append(f" - {name}: {summary}")
return "\n".join(lines)
current = memory.stm.workflow.current or {}
lines = [
f"WORKFLOW SCOPE: active — {current.get('name')} "
f"(stage: {current.get('stage')})",
]
params = current.get("params")
if params:
lines.append(f" Params: {params}")
wf_desc = (workflow.get("description") or "").strip()
if wf_desc:
lines.append(f" Goal: {wf_desc}")
steps = workflow.get("steps", [])
if steps:
lines.append(" Steps:")
for step in steps:
step_id = step.get("id", "?")
step_tool = step.get("tool") or (
"ask_user" if step.get("ask_user") else ""
)
lines.append(f" - {step_id} ({step_tool})")
lines.append(" Call end_workflow(reason) when done, cancelled, or off-topic.")
return "\n".join(lines)
def _format_episodic_context(self, memory) -> str:
"""Format episodic memory context for the prompt."""
lines = []
if memory.episodic.last_search_results:
results = memory.episodic.last_search_results
result_list = results.get("results", [])
lines.append(
f"\nLAST SEARCH: '{results.get('query')}' ({len(result_list)} results)"
)
# Show first 5 results
for i, result in enumerate(result_list[:5]):
name = result.get("name", "Unknown")
lines.append(f" {i + 1}. {name}")
if len(result_list) > 5:
lines.append(f" ... and {len(result_list) - 5} more")
if memory.episodic.pending_question:
question = memory.episodic.pending_question
lines.append(f"\nPENDING QUESTION: {question.get('question')}")
lines.append(f" Type: {question.get('type')}")
if question.get("options"):
lines.append(f" Options: {len(question.get('options'))}")
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)}%")
if memory.episodic.recent_errors:
lines.append("\nRECENT ERRORS (up to 3):")
for error in memory.episodic.recent_errors[-3:]:
lines.append(
f" - Action '{error.get('action')}' failed: {error.get('error')}"
)
# 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 event in unread[:3]:
lines.append(f" - {event.get('type')}: {event.get('data')}")
return "\n".join(lines)
def _format_stm_context(self, memory) -> str:
"""Format short-term memory context for the prompt."""
lines = []
if memory.stm.current_workflow:
workflow = memory.stm.current_workflow
lines.append(
f"CURRENT WORKFLOW: {workflow.get('name')} (stage: {workflow.get('stage')})"
)
if workflow.get("params"):
lines.append(f" Params: {workflow.get('params')}")
if memory.stm.current_topic:
lines.append(f"CURRENT TOPIC: {memory.stm.current_topic}")
if memory.stm.extracted_entities:
lines.append("EXTRACTED ENTITIES:")
for key, value in memory.stm.extracted_entities.items():
lines.append(f" - {key}: {value}")
if memory.stm.language:
lines.append(f"CONVERSATION LANGUAGE: {memory.stm.language}")
return "\n".join(lines)
def _format_memory_schema(self) -> str:
"""Describe available memory components so the agent knows what to read/write and when."""
schema = self._memory_registry.schema()
tier_labels = {
"ltm": "LONG-TERM (persisted)",
"stm": "SHORT-TERM (session)",
"episodic": "EPISODIC (volatile)",
}
lines = ["MEMORY COMPONENTS:"]
for tier, components in schema.items():
if not components:
continue
lines.append(f"\n [{tier_labels.get(tier, tier.upper())}]")
for c in components:
access = c.get("access", "read")
lines.append(f" {c['name']} ({access}): {c['description']}")
for field_name, field_desc in c.get("fields", {}).items():
lines.append(f" · {field_name}: {field_desc}")
return "\n".join(lines)
def _format_config_context(self, memory) -> str:
"""Format configuration context."""
lines = ["CURRENT CONFIGURATION:"]
folders = {
**memory.ltm.workspace.as_dict(),
**memory.ltm.library_paths.to_dict(),
}
if folders:
for key, value in folders.items():
lines.append(f" - {key}: {value}")
else:
lines.append(" (no configuration set)")
return "\n".join(lines)
def build_system_prompt(self) -> str:
"""Build the complete system prompt."""
memory = get_memory()
# Identity + personality
identity = self._format_identity(memory)
# Language instruction
language_instruction = (
"Si la langue de l'user est différente de la langue courante en STM, "
"appelle `set_language` en premier avant de répondre."
)
# Configuration
config_section = self._format_config_context(memory)
# STM context
stm_context = self._format_stm_context(memory)
# Episodic context
episodic_context = self._format_episodic_context(memory)
# Memory schema
memory_schema = self._format_memory_schema()
# Workflow scope (active workflow plan or list of options)
workflow_section = self._format_workflow_scope(memory)
# Available tools (already filtered by scope)
tools_desc = self._format_tools_description()
tools_section = f"\nOUTILS DISPONIBLES:\n{tools_desc}" if tools_desc else ""
rules = """
RÈGLES:
- Utilise les outils pour accomplir les tâches, pas pour décorer
- Si des résultats de recherche sont dispo en mémoire épisodique, référence-les par index
- Confirme toujours avant une opération destructive (move, delete, overwrite)
- Réponses courtes — si c'est fait, dis-le en une ligne
- Si la demande est floue, demande un éclaircissement AVANT de lancer quoi que ce soit
"""
sections = [
identity,
language_instruction,
config_section,
stm_context,
episodic_context,
memory_schema,
workflow_section,
tools_section,
rules,
]
return "\n\n".join(s for s in sections if s and s.strip())