wehub-resource-sync
53 lines
2.1 KiB
Python
53 lines
2.1 KiB
Python
"""Declaratively define the travel-concierge agent in Agent Spec and serialize it."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import os
|
|
|
|
from pyagentspec.agent import Agent
|
|
from pyagentspec.llms import OpenAiCompatibleConfig
|
|
from pyagentspec.serialization import AgentSpecSerializer
|
|
|
|
from .tools import TOOLS
|
|
|
|
SYSTEM_PROMPT = """You are a personal flight concierge with long-term memory.
|
|
|
|
Start every conversation by calling recall_memory to retrieve the traveler's durable
|
|
preferences — home airport, seat preference (e.g. aisle), meal preference (e.g.
|
|
vegetarian), and favorite airlines or destinations. Weave these in naturally; never
|
|
say "according to my memory" or reveal that you looked something up.
|
|
|
|
Use search_flights to find available flights for the requested destination. Present
|
|
the results clearly, highlighting options that best honor the traveler's recalled
|
|
preferences (e.g. nonstop if they prefer it, aisle seat availability, vegetarian meal
|
|
on request, preferred airline or home airport as origin).
|
|
|
|
When the traveler has chosen a flight and wants to book it, call book_flight with the
|
|
chosen flight's id (e.g. 'AMS-001'). The traveler will confirm the booking in the UI
|
|
before it is finalized — do not ask them to confirm again in chat."""
|
|
|
|
|
|
def build_agent() -> Agent:
|
|
# Do NOT set api_key here: AgentSpecSerializer treats it as a SensitiveField and
|
|
# disaggregates it out of the serialized JSON, which then fails to load without a
|
|
# components_registry ("references ... missing ... api_key"). Leave it unset — the
|
|
# LangGraph ChatOpenAI reads OPENAI_API_KEY from the environment (load_dotenv in
|
|
# server.py puts it there).
|
|
llm = OpenAiCompatibleConfig(
|
|
name="concierge_llm",
|
|
model_id=os.getenv("CHAT_MODEL", "gpt-5.4-mini"),
|
|
url=os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1"),
|
|
)
|
|
return Agent(
|
|
name="travel_concierge",
|
|
llm_config=llm,
|
|
system_prompt=SYSTEM_PROMPT,
|
|
tools=TOOLS,
|
|
human_in_the_loop=True,
|
|
)
|
|
|
|
|
|
def build_agent_json() -> str:
|
|
"""Return the Agent Spec JSON string the adapter loads."""
|
|
return AgentSpecSerializer().to_json(build_agent())
|