mirror of
https://github.com/TauricResearch/TradingAgents.git
synced 2026-09-27 06:56:39 +03:00
Yijia-Xiao
cf960d6382
- the Research Manager's plan and the Portfolio Manager's decision live only in investment_plan and final_trade_decision; the debate states no longer copy them as judge_decision - the saved state log writes the Trader's plan as trader_investment_plan, its state key - the report tree, the final display and the live CLI read the decisions from those fields
107 lines
4.5 KiB
Python
107 lines
4.5 KiB
Python
"""Portfolio Manager: synthesises the risk-analyst debate into the final decision.
|
|
|
|
Uses LangChain's ``with_structured_output`` so the LLM produces a typed
|
|
``PortfolioDecision`` directly, in a single call. Its rating is the run's
|
|
``final_rating``, and the decision is rendered to markdown as
|
|
``final_trade_decision`` for the memory log, CLI display and saved reports.
|
|
When a provider does not expose structured output, the agent falls back to
|
|
free-text generation and the rating is read from that text.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from tradingagents.agents.context import (
|
|
get_instrument_context_from_state,
|
|
get_language_instruction,
|
|
get_portfolio_context_from_state,
|
|
)
|
|
from tradingagents.agents.rating import parse_rating
|
|
from tradingagents.agents.schemas import PortfolioDecision, render_pm_decision
|
|
from tradingagents.agents.structured import NO_EXTERNAL_TOOLS, bind_structured, invoke_structured
|
|
|
|
|
|
def create_portfolio_manager(llm):
|
|
structured_llm = bind_structured(llm, PortfolioDecision, "Portfolio Manager")
|
|
|
|
def portfolio_manager_node(state) -> dict:
|
|
instrument_context = get_instrument_context_from_state(state)
|
|
portfolio_context = get_portfolio_context_from_state(state)
|
|
|
|
history = state["risk_debate_state"]["history"]
|
|
risk_debate_state = state["risk_debate_state"]
|
|
research_plan = state["investment_plan"]
|
|
trader_plan = state["trader_investment_plan"]
|
|
|
|
past_context = state.get("past_context", "")
|
|
lessons_line = (
|
|
f"- Lessons from prior decisions and outcomes:\n{past_context}\n"
|
|
if past_context
|
|
else ""
|
|
)
|
|
|
|
prompt = f"""As the Portfolio Manager, synthesize the risk analysts' debate and deliver the final trading decision.
|
|
|
|
{instrument_context}
|
|
|
|
{portfolio_context}
|
|
|
|
---
|
|
|
|
**Rating Scale** (use exactly one):
|
|
- **Buy**: Strong conviction to enter or add to position
|
|
- **Overweight**: Favorable outlook, gradually increase exposure
|
|
- **Hold**: Maintain current position, no action needed
|
|
- **Underweight**: Reduce exposure, take partial profits
|
|
- **Sell**: Exit position or avoid entry
|
|
|
|
**Context:**
|
|
- Research Manager's investment plan: **{research_plan}**
|
|
- Trader's transaction proposal: **{trader_plan}**
|
|
{lessons_line}
|
|
**Risk Analysts Debate History:**
|
|
{history}
|
|
|
|
---
|
|
|
|
Ground every conclusion in specific evidence from the analysts. The risk debate always contains conflicting stances; deciding which is stronger is the job, so conflict alone is not a reason to Hold. Commit to the stronger case, sized by how decisively it wins. Choose Hold only when the evidence is still balanced after that weighing, or too thin to support a call; do not force a direction to appear decisive. Weigh the analysts on their merits, independent of speaking order.
|
|
|
|
## Output
|
|
|
|
Write these sections, in this order, starting with the rating on its own line:
|
|
|
|
- **Rating**: exactly one of Buy / Overweight / Hold / Underweight / Sell
|
|
- **Executive Summary**: the call and how to act on it
|
|
- **Investment Thesis**: the evidence that decided it, and what would change it
|
|
|
|
{NO_EXTERNAL_TOOLS}{get_language_instruction()}"""
|
|
|
|
# The typed rating is the decision; the rendered text only carries it.
|
|
# Read back from text, a rating the thesis quotes could replace it.
|
|
decision = invoke_structured(structured_llm, prompt, "Portfolio Manager")
|
|
if decision is not None:
|
|
final_trade_decision = render_pm_decision(decision)
|
|
final_rating = decision.rating.value
|
|
else:
|
|
final_trade_decision = llm.invoke(prompt).content
|
|
final_rating = parse_rating(final_trade_decision)
|
|
|
|
new_risk_debate_state = {
|
|
"history": risk_debate_state["history"],
|
|
"aggressive_history": risk_debate_state["aggressive_history"],
|
|
"conservative_history": risk_debate_state["conservative_history"],
|
|
"neutral_history": risk_debate_state["neutral_history"],
|
|
"latest_speaker": "Judge",
|
|
"current_aggressive_response": risk_debate_state["current_aggressive_response"],
|
|
"current_conservative_response": risk_debate_state["current_conservative_response"],
|
|
"current_neutral_response": risk_debate_state["current_neutral_response"],
|
|
"count": risk_debate_state["count"],
|
|
}
|
|
|
|
return {
|
|
"risk_debate_state": new_risk_debate_state,
|
|
"final_trade_decision": final_trade_decision,
|
|
"final_rating": final_rating,
|
|
}
|
|
|
|
return portfolio_manager_node
|