mirror of
https://github.com/TauricResearch/TradingAgents.git
synced 2026-09-27 06:56:39 +03:00
Merge remote-tracking branch 'upstream/main' into crypto-analysis-mvp
# Conflicts: # tradingagents/agents/researchers/bear_researcher.py # tradingagents/agents/researchers/bull_researcher.py # tradingagents/graph/propagation.py
This commit is contained in:
CadeYu
@@ -1,2 +0,0 @@
|
||||
import os
|
||||
os.environ.setdefault("PYTHONUTF8", "1")
|
||||
|
||||
@@ -1,6 +1,5 @@
|
||||
from .utils.agent_utils import create_msg_delete
|
||||
from .utils.agent_states import AgentState, InvestDebateState, RiskDebateState
|
||||
from .utils.memory import FinancialSituationMemory
|
||||
|
||||
from .analysts.fundamentals_analyst import create_fundamentals_analyst
|
||||
from .analysts.market_analyst import create_market_analyst
|
||||
@@ -20,7 +19,6 @@ from .managers.portfolio_manager import create_portfolio_manager
|
||||
from .trader.trader import create_trader
|
||||
|
||||
__all__ = [
|
||||
"FinancialSituationMemory",
|
||||
"AgentState",
|
||||
"create_msg_delete",
|
||||
"InvestDebateState",
|
||||
|
||||
@@ -1,26 +1,43 @@
|
||||
from tradingagents.agents.utils.agent_utils import build_instrument_context, get_language_instruction
|
||||
"""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. The result is rendered
|
||||
back to markdown for storage in ``final_trade_decision`` so memory log,
|
||||
CLI display, and saved reports continue to consume the same shape they do
|
||||
today. When a provider does not expose structured output, the agent falls
|
||||
back gracefully to free-text generation.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from tradingagents.agents.schemas import PortfolioDecision, render_pm_decision
|
||||
from tradingagents.agents.utils.agent_utils import (
|
||||
build_instrument_context,
|
||||
get_language_instruction,
|
||||
)
|
||||
from tradingagents.agents.utils.structured import (
|
||||
bind_structured,
|
||||
invoke_structured_or_freetext,
|
||||
)
|
||||
|
||||
|
||||
def create_portfolio_manager(llm, memory):
|
||||
def create_portfolio_manager(llm):
|
||||
structured_llm = bind_structured(llm, PortfolioDecision, "Portfolio Manager")
|
||||
|
||||
def portfolio_manager_node(state) -> dict:
|
||||
|
||||
instrument_context = build_instrument_context(state["company_of_interest"])
|
||||
|
||||
history = state["risk_debate_state"]["history"]
|
||||
risk_debate_state = state["risk_debate_state"]
|
||||
market_research_report = state["market_report"]
|
||||
news_report = state["news_report"]
|
||||
fundamentals_report = state["fundamentals_report"]
|
||||
sentiment_report = state["sentiment_report"]
|
||||
research_plan = state["investment_plan"]
|
||||
trader_plan = state["trader_investment_plan"]
|
||||
|
||||
curr_situation = f"{market_research_report}\n\n{sentiment_report}\n\n{news_report}\n\n{fundamentals_report}"
|
||||
past_memories = memory.get_memories(curr_situation, n_matches=2)
|
||||
|
||||
past_memory_str = ""
|
||||
for i, rec in enumerate(past_memories, 1):
|
||||
past_memory_str += rec["recommendation"] + "\n\n"
|
||||
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.
|
||||
|
||||
@@ -38,15 +55,7 @@ def create_portfolio_manager(llm, memory):
|
||||
**Context:**
|
||||
- Research Manager's investment plan: **{research_plan}**
|
||||
- Trader's transaction proposal: **{trader_plan}**
|
||||
- Lessons from past decisions: **{past_memory_str}**
|
||||
|
||||
**Required Output Structure:**
|
||||
1. **Rating**: State one of Buy / Overweight / Hold / Underweight / Sell.
|
||||
2. **Executive Summary**: A concise action plan covering entry strategy, position sizing, key risk levels, and time horizon.
|
||||
3. **Investment Thesis**: Detailed reasoning anchored in the analysts' debate and past reflections.
|
||||
|
||||
---
|
||||
|
||||
{lessons_line}
|
||||
**Risk Analysts Debate History:**
|
||||
{history}
|
||||
|
||||
@@ -54,10 +63,16 @@ def create_portfolio_manager(llm, memory):
|
||||
|
||||
Be decisive and ground every conclusion in specific evidence from the analysts.{get_language_instruction()}"""
|
||||
|
||||
response = llm.invoke(prompt)
|
||||
final_trade_decision = invoke_structured_or_freetext(
|
||||
structured_llm,
|
||||
llm,
|
||||
prompt,
|
||||
render_pm_decision,
|
||||
"Portfolio Manager",
|
||||
)
|
||||
|
||||
new_risk_debate_state = {
|
||||
"judge_decision": response.content,
|
||||
"judge_decision": final_trade_decision,
|
||||
"history": risk_debate_state["history"],
|
||||
"aggressive_history": risk_debate_state["aggressive_history"],
|
||||
"conservative_history": risk_debate_state["conservative_history"],
|
||||
@@ -71,7 +86,7 @@ Be decisive and ground every conclusion in specific evidence from the analysts.{
|
||||
|
||||
return {
|
||||
"risk_debate_state": new_risk_debate_state,
|
||||
"final_trade_decision": response.content,
|
||||
"final_trade_decision": final_trade_decision,
|
||||
}
|
||||
|
||||
return portfolio_manager_node
|
||||
|
||||
@@ -1,58 +1,64 @@
|
||||
"""Research Manager: turns the bull/bear debate into a structured investment plan for the trader."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from tradingagents.agents.schemas import ResearchPlan, render_research_plan
|
||||
from tradingagents.agents.utils.agent_utils import build_instrument_context
|
||||
from tradingagents.agents.utils.structured import (
|
||||
bind_structured,
|
||||
invoke_structured_or_freetext,
|
||||
)
|
||||
|
||||
|
||||
def create_research_manager(llm, memory):
|
||||
def create_research_manager(llm):
|
||||
structured_llm = bind_structured(llm, ResearchPlan, "Research Manager")
|
||||
|
||||
def research_manager_node(state) -> dict:
|
||||
instrument_context = build_instrument_context(state["company_of_interest"])
|
||||
history = state["investment_debate_state"].get("history", "")
|
||||
market_research_report = state["market_report"]
|
||||
sentiment_report = state["sentiment_report"]
|
||||
news_report = state["news_report"]
|
||||
fundamentals_report = state["fundamentals_report"]
|
||||
|
||||
investment_debate_state = state["investment_debate_state"]
|
||||
|
||||
curr_situation = f"{market_research_report}\n\n{sentiment_report}\n\n{news_report}\n\n{fundamentals_report}"
|
||||
past_memories = memory.get_memories(curr_situation, n_matches=2)
|
||||
|
||||
past_memory_str = ""
|
||||
for i, rec in enumerate(past_memories, 1):
|
||||
past_memory_str += rec["recommendation"] + "\n\n"
|
||||
|
||||
prompt = f"""As the portfolio manager and debate facilitator, your role is to critically evaluate this round of debate and make a definitive decision: align with the bear analyst, the bull analyst, or choose Hold only if it is strongly justified based on the arguments presented.
|
||||
|
||||
Summarize the key points from both sides concisely, focusing on the most compelling evidence or reasoning. Your recommendation—Buy, Sell, or Hold—must be clear and actionable. Avoid defaulting to Hold simply because both sides have valid points; commit to a stance grounded in the debate's strongest arguments.
|
||||
|
||||
Additionally, develop a detailed investment plan for the trader. This should include:
|
||||
|
||||
Your Recommendation: A decisive stance supported by the most convincing arguments.
|
||||
Rationale: An explanation of why these arguments lead to your conclusion.
|
||||
Strategic Actions: Concrete steps for implementing the recommendation.
|
||||
Take into account your past mistakes on similar situations. Use these insights to refine your decision-making and ensure you are learning and improving. Present your analysis conversationally, as if speaking naturally, without special formatting.
|
||||
|
||||
Here are your past reflections on mistakes:
|
||||
\"{past_memory_str}\"
|
||||
prompt = f"""As the Research Manager and debate facilitator, your role is to critically evaluate this round of debate and deliver a clear, actionable investment plan for the trader.
|
||||
|
||||
{instrument_context}
|
||||
|
||||
Here is the debate:
|
||||
Debate History:
|
||||
---
|
||||
|
||||
**Rating Scale** (use exactly one):
|
||||
- **Buy**: Strong conviction in the bull thesis; recommend taking or growing the position
|
||||
- **Overweight**: Constructive view; recommend gradually increasing exposure
|
||||
- **Hold**: Balanced view; recommend maintaining the current position
|
||||
- **Underweight**: Cautious view; recommend trimming exposure
|
||||
- **Sell**: Strong conviction in the bear thesis; recommend exiting or avoiding the position
|
||||
|
||||
Commit to a clear stance whenever the debate's strongest arguments warrant one; reserve Hold for situations where the evidence on both sides is genuinely balanced.
|
||||
|
||||
---
|
||||
|
||||
**Debate History:**
|
||||
{history}"""
|
||||
response = llm.invoke(prompt)
|
||||
|
||||
investment_plan = invoke_structured_or_freetext(
|
||||
structured_llm,
|
||||
llm,
|
||||
prompt,
|
||||
render_research_plan,
|
||||
"Research Manager",
|
||||
)
|
||||
|
||||
new_investment_debate_state = {
|
||||
"judge_decision": response.content,
|
||||
"judge_decision": investment_plan,
|
||||
"history": investment_debate_state.get("history", ""),
|
||||
"bear_history": investment_debate_state.get("bear_history", ""),
|
||||
"bull_history": investment_debate_state.get("bull_history", ""),
|
||||
"current_response": response.content,
|
||||
"current_response": investment_plan,
|
||||
"count": investment_debate_state["count"],
|
||||
}
|
||||
|
||||
return {
|
||||
"investment_debate_state": new_investment_debate_state,
|
||||
"investment_plan": response.content,
|
||||
"investment_plan": investment_plan,
|
||||
}
|
||||
|
||||
return research_manager_node
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
|
||||
|
||||
def create_bear_researcher(llm, memory):
|
||||
def create_bear_researcher(llm):
|
||||
def bear_node(state) -> dict:
|
||||
investment_debate_state = state["investment_debate_state"]
|
||||
history = investment_debate_state.get("history", "")
|
||||
@@ -19,13 +19,6 @@ def create_bear_researcher(llm, memory):
|
||||
else "Asset fundamentals report (may be unavailable for crypto)"
|
||||
)
|
||||
|
||||
curr_situation = f"{market_research_report}\n\n{sentiment_report}\n\n{news_report}\n\n{fundamentals_report}"
|
||||
past_memories = memory.get_memories(curr_situation, n_matches=2)
|
||||
|
||||
past_memory_str = ""
|
||||
for i, rec in enumerate(past_memories, 1):
|
||||
past_memory_str += rec["recommendation"] + "\n\n"
|
||||
|
||||
prompt = f"""You are a Bear Analyst making the case against investing in the {target_label}. Your goal is to present a well-reasoned argument emphasizing risks, challenges, and negative indicators. Leverage the provided research and data to highlight potential downsides and counter bullish arguments effectively.
|
||||
|
||||
Key points to focus on:
|
||||
@@ -44,8 +37,7 @@ Latest world affairs news: {news_report}
|
||||
{fundamentals_label}: {fundamentals_report}
|
||||
Conversation history of the debate: {history}
|
||||
Last bull argument: {current_response}
|
||||
Reflections from similar situations and lessons learned: {past_memory_str}
|
||||
Use this information to deliver a compelling bear argument, refute the bull's claims, and engage in a dynamic debate that demonstrates the risks and weaknesses of investing in the {target_label}. You must also address reflections and learn from lessons and mistakes you made in the past.
|
||||
Use this information to deliver a compelling bear argument, refute the bull's claims, and engage in a dynamic debate that demonstrates the risks and weaknesses of investing in the {target_label}.
|
||||
"""
|
||||
|
||||
response = llm.invoke(prompt)
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
|
||||
|
||||
def create_bull_researcher(llm, memory):
|
||||
def create_bull_researcher(llm):
|
||||
def bull_node(state) -> dict:
|
||||
investment_debate_state = state["investment_debate_state"]
|
||||
history = investment_debate_state.get("history", "")
|
||||
@@ -19,13 +19,6 @@ def create_bull_researcher(llm, memory):
|
||||
else "Asset fundamentals report (may be unavailable for crypto)"
|
||||
)
|
||||
|
||||
curr_situation = f"{market_research_report}\n\n{sentiment_report}\n\n{news_report}\n\n{fundamentals_report}"
|
||||
past_memories = memory.get_memories(curr_situation, n_matches=2)
|
||||
|
||||
past_memory_str = ""
|
||||
for i, rec in enumerate(past_memories, 1):
|
||||
past_memory_str += rec["recommendation"] + "\n\n"
|
||||
|
||||
prompt = f"""You are a Bull Analyst advocating for investing in the {target_label}. Your task is to build a strong, evidence-based case emphasizing growth potential, competitive advantages, and positive market indicators. Leverage the provided research and data to address concerns and counter bearish arguments effectively.
|
||||
|
||||
Key points to focus on:
|
||||
@@ -42,8 +35,7 @@ Latest world affairs news: {news_report}
|
||||
{fundamentals_label}: {fundamentals_report}
|
||||
Conversation history of the debate: {history}
|
||||
Last bear argument: {current_response}
|
||||
Reflections from similar situations and lessons learned: {past_memory_str}
|
||||
Use this information to deliver a compelling bull argument, refute the bear's concerns, and engage in a dynamic debate that demonstrates the strengths of the bull position. You must also address reflections and learn from lessons and mistakes you made in the past.
|
||||
Use this information to deliver a compelling bull argument, refute the bear's concerns, and engage in a dynamic debate that demonstrates the strengths of the bull position.
|
||||
"""
|
||||
|
||||
response = llm.invoke(prompt)
|
||||
|
||||
@@ -0,0 +1,228 @@
|
||||
"""Pydantic schemas used by agents that produce structured output.
|
||||
|
||||
The framework's primary artifact is still prose: each agent's natural-language
|
||||
reasoning is what users read in the saved markdown reports and what the
|
||||
downstream agents read as context. Structured output is layered onto the
|
||||
three decision-making agents (Research Manager, Trader, Portfolio Manager)
|
||||
so that:
|
||||
|
||||
- Their outputs follow consistent section headers across runs and providers
|
||||
- Each provider's native structured-output mode is used (json_schema for
|
||||
OpenAI/xAI, response_schema for Gemini, tool-use for Anthropic)
|
||||
- Schema field descriptions become the model's output instructions, freeing
|
||||
the prompt body to focus on context and the rating-scale guidance
|
||||
- A render helper turns the parsed Pydantic instance back into the same
|
||||
markdown shape the rest of the system already consumes, so display,
|
||||
memory log, and saved reports keep working unchanged
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from enum import Enum
|
||||
from typing import Optional
|
||||
|
||||
from pydantic import BaseModel, Field
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Shared rating types
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class PortfolioRating(str, Enum):
|
||||
"""5-tier rating used by the Research Manager and Portfolio Manager."""
|
||||
|
||||
BUY = "Buy"
|
||||
OVERWEIGHT = "Overweight"
|
||||
HOLD = "Hold"
|
||||
UNDERWEIGHT = "Underweight"
|
||||
SELL = "Sell"
|
||||
|
||||
|
||||
class TraderAction(str, Enum):
|
||||
"""3-tier transaction direction used by the Trader.
|
||||
|
||||
The Trader's job is to translate the Research Manager's investment plan
|
||||
into a concrete transaction proposal: should the desk execute a Buy, a
|
||||
Sell, or sit on Hold this round. Position sizing and the nuanced
|
||||
Overweight / Underweight calls happen later at the Portfolio Manager.
|
||||
"""
|
||||
|
||||
BUY = "Buy"
|
||||
HOLD = "Hold"
|
||||
SELL = "Sell"
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Research Manager
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class ResearchPlan(BaseModel):
|
||||
"""Structured investment plan produced by the Research Manager.
|
||||
|
||||
Hand-off to the Trader: the recommendation pins the directional view,
|
||||
the rationale captures which side of the bull/bear debate carried the
|
||||
argument, and the strategic actions translate that into concrete
|
||||
instructions the trader can execute against.
|
||||
"""
|
||||
|
||||
recommendation: PortfolioRating = Field(
|
||||
description=(
|
||||
"The investment recommendation. Exactly one of Buy / Overweight / "
|
||||
"Hold / Underweight / Sell. Reserve Hold for situations where the "
|
||||
"evidence on both sides is genuinely balanced; otherwise commit to "
|
||||
"the side with the stronger arguments."
|
||||
),
|
||||
)
|
||||
rationale: str = Field(
|
||||
description=(
|
||||
"Conversational summary of the key points from both sides of the "
|
||||
"debate, ending with which arguments led to the recommendation. "
|
||||
"Speak naturally, as if to a teammate."
|
||||
),
|
||||
)
|
||||
strategic_actions: str = Field(
|
||||
description=(
|
||||
"Concrete steps for the trader to implement the recommendation, "
|
||||
"including position sizing guidance consistent with the rating."
|
||||
),
|
||||
)
|
||||
|
||||
|
||||
def render_research_plan(plan: ResearchPlan) -> str:
|
||||
"""Render a ResearchPlan to markdown for storage and the trader's prompt context."""
|
||||
return "\n".join([
|
||||
f"**Recommendation**: {plan.recommendation.value}",
|
||||
"",
|
||||
f"**Rationale**: {plan.rationale}",
|
||||
"",
|
||||
f"**Strategic Actions**: {plan.strategic_actions}",
|
||||
])
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Trader
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class TraderProposal(BaseModel):
|
||||
"""Structured transaction proposal produced by the Trader.
|
||||
|
||||
The trader reads the Research Manager's investment plan and the analyst
|
||||
reports, then turns them into a concrete transaction: what action to
|
||||
take, the reasoning that justifies it, and the practical levels for
|
||||
entry, stop-loss, and sizing.
|
||||
"""
|
||||
|
||||
action: TraderAction = Field(
|
||||
description="The transaction direction. Exactly one of Buy / Hold / Sell.",
|
||||
)
|
||||
reasoning: str = Field(
|
||||
description=(
|
||||
"The case for this action, anchored in the analysts' reports and "
|
||||
"the research plan. Two to four sentences."
|
||||
),
|
||||
)
|
||||
entry_price: Optional[float] = Field(
|
||||
default=None,
|
||||
description="Optional entry price target in the instrument's quote currency.",
|
||||
)
|
||||
stop_loss: Optional[float] = Field(
|
||||
default=None,
|
||||
description="Optional stop-loss price in the instrument's quote currency.",
|
||||
)
|
||||
position_sizing: Optional[str] = Field(
|
||||
default=None,
|
||||
description="Optional sizing guidance, e.g. '5% of portfolio'.",
|
||||
)
|
||||
|
||||
|
||||
def render_trader_proposal(proposal: TraderProposal) -> str:
|
||||
"""Render a TraderProposal to markdown.
|
||||
|
||||
The trailing ``FINAL TRANSACTION PROPOSAL: **BUY/HOLD/SELL**`` line is
|
||||
preserved for backward compatibility with the analyst stop-signal text
|
||||
and any external code that greps for it.
|
||||
"""
|
||||
parts = [
|
||||
f"**Action**: {proposal.action.value}",
|
||||
"",
|
||||
f"**Reasoning**: {proposal.reasoning}",
|
||||
]
|
||||
if proposal.entry_price is not None:
|
||||
parts.extend(["", f"**Entry Price**: {proposal.entry_price}"])
|
||||
if proposal.stop_loss is not None:
|
||||
parts.extend(["", f"**Stop Loss**: {proposal.stop_loss}"])
|
||||
if proposal.position_sizing:
|
||||
parts.extend(["", f"**Position Sizing**: {proposal.position_sizing}"])
|
||||
parts.extend([
|
||||
"",
|
||||
f"FINAL TRANSACTION PROPOSAL: **{proposal.action.value.upper()}**",
|
||||
])
|
||||
return "\n".join(parts)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Portfolio Manager
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class PortfolioDecision(BaseModel):
|
||||
"""Structured output produced by the Portfolio Manager.
|
||||
|
||||
The model fills every field as part of its primary LLM call; no separate
|
||||
extraction pass is required. Field descriptions double as the model's
|
||||
output instructions, so the prompt body only needs to convey context and
|
||||
the rating-scale guidance.
|
||||
"""
|
||||
|
||||
rating: PortfolioRating = Field(
|
||||
description=(
|
||||
"The final position rating. Exactly one of Buy / Overweight / Hold / "
|
||||
"Underweight / Sell, picked based on the analysts' debate."
|
||||
),
|
||||
)
|
||||
executive_summary: str = Field(
|
||||
description=(
|
||||
"A concise action plan covering entry strategy, position sizing, "
|
||||
"key risk levels, and time horizon. Two to four sentences."
|
||||
),
|
||||
)
|
||||
investment_thesis: str = Field(
|
||||
description=(
|
||||
"Detailed reasoning anchored in specific evidence from the analysts' "
|
||||
"debate. If prior lessons are referenced in the prompt context, "
|
||||
"incorporate them; otherwise rely solely on the current analysis."
|
||||
),
|
||||
)
|
||||
price_target: Optional[float] = Field(
|
||||
default=None,
|
||||
description="Optional target price in the instrument's quote currency.",
|
||||
)
|
||||
time_horizon: Optional[str] = Field(
|
||||
default=None,
|
||||
description="Optional recommended holding period, e.g. '3-6 months'.",
|
||||
)
|
||||
|
||||
|
||||
def render_pm_decision(decision: PortfolioDecision) -> str:
|
||||
"""Render a PortfolioDecision back to the markdown shape the rest of the system expects.
|
||||
|
||||
Memory log, CLI display, and saved report files all read this markdown,
|
||||
so the rendered output preserves the exact section headers (``**Rating**``,
|
||||
``**Executive Summary**``, ``**Investment Thesis**``) that downstream
|
||||
parsers and the report writers already handle.
|
||||
"""
|
||||
parts = [
|
||||
f"**Rating**: {decision.rating.value}",
|
||||
"",
|
||||
f"**Executive Summary**: {decision.executive_summary}",
|
||||
"",
|
||||
f"**Investment Thesis**: {decision.investment_thesis}",
|
||||
]
|
||||
if decision.price_target is not None:
|
||||
parts.extend(["", f"**Price Target**: {decision.price_target}"])
|
||||
if decision.time_horizon:
|
||||
parts.extend(["", f"**Time Horizon**: {decision.time_horizon}"])
|
||||
return "\n".join(parts)
|
||||
@@ -1,47 +1,61 @@
|
||||
"""Trader: turns the Research Manager's investment plan into a concrete transaction proposal."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import functools
|
||||
|
||||
from langchain_core.messages import AIMessage
|
||||
|
||||
from tradingagents.agents.schemas import TraderProposal, render_trader_proposal
|
||||
from tradingagents.agents.utils.agent_utils import build_instrument_context
|
||||
from tradingagents.agents.utils.structured import (
|
||||
bind_structured,
|
||||
invoke_structured_or_freetext,
|
||||
)
|
||||
|
||||
|
||||
def create_trader(llm, memory):
|
||||
def create_trader(llm):
|
||||
structured_llm = bind_structured(llm, TraderProposal, "Trader")
|
||||
|
||||
def trader_node(state, name):
|
||||
company_name = state["company_of_interest"]
|
||||
asset_type = state.get("asset_type", "stock")
|
||||
instrument_context = build_instrument_context(company_name, asset_type)
|
||||
investment_plan = state["investment_plan"]
|
||||
market_research_report = state["market_report"]
|
||||
sentiment_report = state["sentiment_report"]
|
||||
news_report = state["news_report"]
|
||||
fundamentals_report = state["fundamentals_report"]
|
||||
|
||||
curr_situation = f"{market_research_report}\n\n{sentiment_report}\n\n{news_report}\n\n{fundamentals_report}"
|
||||
past_memories = memory.get_memories(curr_situation, n_matches=2)
|
||||
|
||||
past_memory_str = ""
|
||||
if past_memories:
|
||||
for i, rec in enumerate(past_memories, 1):
|
||||
past_memory_str += rec["recommendation"] + "\n\n"
|
||||
else:
|
||||
past_memory_str = "No past memories found."
|
||||
|
||||
context = {
|
||||
"role": "user",
|
||||
"content": f"Based on a comprehensive analysis by a team of analysts, here is an investment plan tailored for {company_name}. {instrument_context} This plan incorporates insights from current technical market trends, macroeconomic indicators, and social media sentiment. Use this plan as a foundation for evaluating your next trading decision.\n\nProposed Investment Plan: {investment_plan}\n\nLeverage these insights to make an informed and strategic decision.",
|
||||
}
|
||||
|
||||
messages = [
|
||||
{
|
||||
"role": "system",
|
||||
"content": f"""You are a trading agent analyzing market data to make investment decisions. Based on your analysis, provide a specific recommendation to buy, sell, or hold. End with a firm decision and always conclude your response with 'FINAL TRANSACTION PROPOSAL: **BUY/HOLD/SELL**' to confirm your recommendation. Apply lessons from past decisions to strengthen your analysis. Here are reflections from similar situations you traded in and the lessons learned: {past_memory_str}""",
|
||||
"content": (
|
||||
"You are a trading agent analyzing market data to make investment decisions. "
|
||||
"Based on your analysis, provide a specific recommendation to buy, sell, or hold. "
|
||||
"Anchor your reasoning in the analysts' reports and the research plan."
|
||||
),
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": (
|
||||
f"Based on a comprehensive analysis by a team of analysts, here is an investment "
|
||||
f"plan tailored for {company_name}. {instrument_context} This plan incorporates "
|
||||
f"insights from current technical market trends, macroeconomic indicators, and "
|
||||
f"social media sentiment. Use this plan as a foundation for evaluating your next "
|
||||
f"trading decision.\n\nProposed Investment Plan: {investment_plan}\n\n"
|
||||
f"Leverage these insights to make an informed and strategic decision."
|
||||
),
|
||||
},
|
||||
context,
|
||||
]
|
||||
|
||||
result = llm.invoke(messages)
|
||||
trader_plan = invoke_structured_or_freetext(
|
||||
structured_llm,
|
||||
llm,
|
||||
messages,
|
||||
render_trader_proposal,
|
||||
"Trader",
|
||||
)
|
||||
|
||||
return {
|
||||
"messages": [result],
|
||||
"trader_investment_plan": result.content,
|
||||
"messages": [AIMessage(content=trader_plan)],
|
||||
"trader_investment_plan": trader_plan,
|
||||
"sender": name,
|
||||
}
|
||||
|
||||
|
||||
@@ -71,3 +71,4 @@ class AgentState(MessagesState):
|
||||
RiskDebateState, "Current state of the debate on evaluating risk"
|
||||
]
|
||||
final_trade_decision: Annotated[str, "Final decision made by the Risk Analysts"]
|
||||
past_context: Annotated[str, "Memory log context injected at run start (same-ticker decisions + cross-ticker lessons)"]
|
||||
|
||||
@@ -1,144 +1,300 @@
|
||||
"""Financial situation memory using BM25 for lexical similarity matching.
|
||||
"""Append-only markdown decision log for TradingAgents."""
|
||||
|
||||
Uses BM25 (Best Matching 25) algorithm for retrieval - no API calls,
|
||||
no token limits, works offline with any LLM provider.
|
||||
"""
|
||||
|
||||
from rank_bm25 import BM25Okapi
|
||||
from typing import List, Tuple
|
||||
from typing import List, Optional
|
||||
from pathlib import Path
|
||||
import re
|
||||
|
||||
from tradingagents.agents.utils.rating import parse_rating
|
||||
|
||||
class FinancialSituationMemory:
|
||||
"""Memory system for storing and retrieving financial situations using BM25."""
|
||||
|
||||
def __init__(self, name: str, config: dict = None):
|
||||
"""Initialize the memory system.
|
||||
class TradingMemoryLog:
|
||||
"""Append-only markdown log of trading decisions and reflections."""
|
||||
|
||||
Args:
|
||||
name: Name identifier for this memory instance
|
||||
config: Configuration dict (kept for API compatibility, not used for BM25)
|
||||
"""
|
||||
self.name = name
|
||||
self.documents: List[str] = []
|
||||
self.recommendations: List[str] = []
|
||||
self.bm25 = None
|
||||
# HTML comment: cannot appear in LLM prose output, safe as a hard delimiter
|
||||
_SEPARATOR = "\n\n<!-- ENTRY_END -->\n\n"
|
||||
# Precompiled patterns — avoids re-compilation on every load_entries() call
|
||||
_DECISION_RE = re.compile(r"DECISION:\n(.*?)(?=\nREFLECTION:|\Z)", re.DOTALL)
|
||||
_REFLECTION_RE = re.compile(r"REFLECTION:\n(.*?)$", re.DOTALL)
|
||||
|
||||
def _tokenize(self, text: str) -> List[str]:
|
||||
"""Tokenize text for BM25 indexing.
|
||||
def __init__(self, config: dict = None):
|
||||
cfg = config or {}
|
||||
self._log_path = None
|
||||
path = cfg.get("memory_log_path")
|
||||
if path:
|
||||
self._log_path = Path(path).expanduser()
|
||||
self._log_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
# Optional cap on resolved entries. None disables rotation.
|
||||
self._max_entries = cfg.get("memory_log_max_entries")
|
||||
|
||||
Simple whitespace + punctuation tokenization with lowercasing.
|
||||
"""
|
||||
# Lowercase and split on non-alphanumeric characters
|
||||
tokens = re.findall(r'\b\w+\b', text.lower())
|
||||
return tokens
|
||||
# --- Write path (Phase A) ---
|
||||
|
||||
def _rebuild_index(self):
|
||||
"""Rebuild the BM25 index after adding documents."""
|
||||
if self.documents:
|
||||
tokenized_docs = [self._tokenize(doc) for doc in self.documents]
|
||||
self.bm25 = BM25Okapi(tokenized_docs)
|
||||
else:
|
||||
self.bm25 = None
|
||||
def store_decision(
|
||||
self,
|
||||
ticker: str,
|
||||
trade_date: str,
|
||||
final_trade_decision: str,
|
||||
) -> None:
|
||||
"""Append pending entry at end of propagate(). No LLM call."""
|
||||
if not self._log_path:
|
||||
return
|
||||
# Idempotency guard: fast raw-text scan instead of full parse
|
||||
if self._log_path.exists():
|
||||
raw = self._log_path.read_text(encoding="utf-8")
|
||||
for line in raw.splitlines():
|
||||
if line.startswith(f"[{trade_date} | {ticker} |") and line.endswith("| pending]"):
|
||||
return
|
||||
rating = parse_rating(final_trade_decision)
|
||||
tag = f"[{trade_date} | {ticker} | {rating} | pending]"
|
||||
entry = f"{tag}\n\nDECISION:\n{final_trade_decision}{self._SEPARATOR}"
|
||||
with open(self._log_path, "a", encoding="utf-8") as f:
|
||||
f.write(entry)
|
||||
|
||||
def add_situations(self, situations_and_advice: List[Tuple[str, str]]):
|
||||
"""Add financial situations and their corresponding advice.
|
||||
# --- Read path (Phase A) ---
|
||||
|
||||
Args:
|
||||
situations_and_advice: List of tuples (situation, recommendation)
|
||||
"""
|
||||
for situation, recommendation in situations_and_advice:
|
||||
self.documents.append(situation)
|
||||
self.recommendations.append(recommendation)
|
||||
|
||||
# Rebuild BM25 index with new documents
|
||||
self._rebuild_index()
|
||||
|
||||
def get_memories(self, current_situation: str, n_matches: int = 1) -> List[dict]:
|
||||
"""Find matching recommendations using BM25 similarity.
|
||||
|
||||
Args:
|
||||
current_situation: The current financial situation to match against
|
||||
n_matches: Number of top matches to return
|
||||
|
||||
Returns:
|
||||
List of dicts with matched_situation, recommendation, and similarity_score
|
||||
"""
|
||||
if not self.documents or self.bm25 is None:
|
||||
def load_entries(self) -> List[dict]:
|
||||
"""Parse all entries from log. Returns list of dicts."""
|
||||
if not self._log_path or not self._log_path.exists():
|
||||
return []
|
||||
text = self._log_path.read_text(encoding="utf-8")
|
||||
raw_entries = [e.strip() for e in text.split(self._SEPARATOR) if e.strip()]
|
||||
entries = []
|
||||
for raw in raw_entries:
|
||||
parsed = self._parse_entry(raw)
|
||||
if parsed:
|
||||
entries.append(parsed)
|
||||
return entries
|
||||
|
||||
# Tokenize query
|
||||
query_tokens = self._tokenize(current_situation)
|
||||
def get_pending_entries(self) -> List[dict]:
|
||||
"""Return entries with outcome:pending (for Phase B)."""
|
||||
return [e for e in self.load_entries() if e.get("pending")]
|
||||
|
||||
# Get BM25 scores for all documents
|
||||
scores = self.bm25.get_scores(query_tokens)
|
||||
def get_past_context(self, ticker: str, n_same: int = 5, n_cross: int = 3) -> str:
|
||||
"""Return formatted past context string for agent prompt injection."""
|
||||
entries = [e for e in self.load_entries() if not e.get("pending")]
|
||||
if not entries:
|
||||
return ""
|
||||
|
||||
# Get top-n indices sorted by score (descending)
|
||||
top_indices = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)[:n_matches]
|
||||
same, cross = [], []
|
||||
for e in reversed(entries):
|
||||
if len(same) >= n_same and len(cross) >= n_cross:
|
||||
break
|
||||
if e["ticker"] == ticker and len(same) < n_same:
|
||||
same.append(e)
|
||||
elif e["ticker"] != ticker and len(cross) < n_cross:
|
||||
cross.append(e)
|
||||
|
||||
# Build results
|
||||
results = []
|
||||
max_score = float(scores.max()) if len(scores) > 0 and scores.max() > 0 else 1.0
|
||||
if not same and not cross:
|
||||
return ""
|
||||
|
||||
for idx in top_indices:
|
||||
# Normalize score to 0-1 range for consistency
|
||||
normalized_score = scores[idx] / max_score if max_score > 0 else 0
|
||||
results.append({
|
||||
"matched_situation": self.documents[idx],
|
||||
"recommendation": self.recommendations[idx],
|
||||
"similarity_score": normalized_score,
|
||||
})
|
||||
parts = []
|
||||
if same:
|
||||
parts.append(f"Past analyses of {ticker} (most recent first):")
|
||||
parts.extend(self._format_full(e) for e in same)
|
||||
if cross:
|
||||
parts.append("Recent cross-ticker lessons:")
|
||||
parts.extend(self._format_reflection_only(e) for e in cross)
|
||||
return "\n\n".join(parts)
|
||||
|
||||
return results
|
||||
# --- Update path (Phase B) ---
|
||||
|
||||
def clear(self):
|
||||
"""Clear all stored memories."""
|
||||
self.documents = []
|
||||
self.recommendations = []
|
||||
self.bm25 = None
|
||||
def update_with_outcome(
|
||||
self,
|
||||
ticker: str,
|
||||
trade_date: str,
|
||||
raw_return: float,
|
||||
alpha_return: float,
|
||||
holding_days: int,
|
||||
reflection: str,
|
||||
) -> None:
|
||||
"""Replace pending tag and append REFLECTION section using atomic write.
|
||||
|
||||
Finds the first pending entry matching (trade_date, ticker), updates
|
||||
its tag with return figures, and appends a REFLECTION section. Uses
|
||||
a temp-file + os.replace() so a crash mid-write never corrupts the log.
|
||||
"""
|
||||
if not self._log_path or not self._log_path.exists():
|
||||
return
|
||||
|
||||
if __name__ == "__main__":
|
||||
# Example usage
|
||||
matcher = FinancialSituationMemory("test_memory")
|
||||
text = self._log_path.read_text(encoding="utf-8")
|
||||
blocks = text.split(self._SEPARATOR)
|
||||
|
||||
# Example data
|
||||
example_data = [
|
||||
(
|
||||
"High inflation rate with rising interest rates and declining consumer spending",
|
||||
"Consider defensive sectors like consumer staples and utilities. Review fixed-income portfolio duration.",
|
||||
),
|
||||
(
|
||||
"Tech sector showing high volatility with increasing institutional selling pressure",
|
||||
"Reduce exposure to high-growth tech stocks. Look for value opportunities in established tech companies with strong cash flows.",
|
||||
),
|
||||
(
|
||||
"Strong dollar affecting emerging markets with increasing forex volatility",
|
||||
"Hedge currency exposure in international positions. Consider reducing allocation to emerging market debt.",
|
||||
),
|
||||
(
|
||||
"Market showing signs of sector rotation with rising yields",
|
||||
"Rebalance portfolio to maintain target allocations. Consider increasing exposure to sectors benefiting from higher rates.",
|
||||
),
|
||||
]
|
||||
pending_prefix = f"[{trade_date} | {ticker} |"
|
||||
raw_pct = f"{raw_return:+.1%}"
|
||||
alpha_pct = f"{alpha_return:+.1%}"
|
||||
|
||||
# Add the example situations and recommendations
|
||||
matcher.add_situations(example_data)
|
||||
updated = False
|
||||
new_blocks = []
|
||||
for block in blocks:
|
||||
stripped = block.strip()
|
||||
if not stripped:
|
||||
new_blocks.append(block)
|
||||
continue
|
||||
|
||||
# Example query
|
||||
current_situation = """
|
||||
Market showing increased volatility in tech sector, with institutional investors
|
||||
reducing positions and rising interest rates affecting growth stock valuations
|
||||
"""
|
||||
lines = stripped.splitlines()
|
||||
tag_line = lines[0].strip()
|
||||
|
||||
try:
|
||||
recommendations = matcher.get_memories(current_situation, n_matches=2)
|
||||
if (
|
||||
not updated
|
||||
and tag_line.startswith(pending_prefix)
|
||||
and tag_line.endswith("| pending]")
|
||||
):
|
||||
# Parse rating from the existing pending tag
|
||||
fields = [f.strip() for f in tag_line[1:-1].split("|")]
|
||||
rating = fields[2]
|
||||
new_tag = (
|
||||
f"[{trade_date} | {ticker} | {rating}"
|
||||
f" | {raw_pct} | {alpha_pct} | {holding_days}d]"
|
||||
)
|
||||
rest = "\n".join(lines[1:])
|
||||
new_blocks.append(
|
||||
f"{new_tag}\n\n{rest.lstrip()}\n\nREFLECTION:\n{reflection}"
|
||||
)
|
||||
updated = True
|
||||
else:
|
||||
new_blocks.append(block)
|
||||
|
||||
for i, rec in enumerate(recommendations, 1):
|
||||
print(f"\nMatch {i}:")
|
||||
print(f"Similarity Score: {rec['similarity_score']:.2f}")
|
||||
print(f"Matched Situation: {rec['matched_situation']}")
|
||||
print(f"Recommendation: {rec['recommendation']}")
|
||||
if not updated:
|
||||
return
|
||||
|
||||
except Exception as e:
|
||||
print(f"Error during recommendation: {str(e)}")
|
||||
new_blocks = self._apply_rotation(new_blocks)
|
||||
new_text = self._SEPARATOR.join(new_blocks)
|
||||
tmp_path = self._log_path.with_suffix(".tmp")
|
||||
tmp_path.write_text(new_text, encoding="utf-8")
|
||||
tmp_path.replace(self._log_path)
|
||||
|
||||
def batch_update_with_outcomes(self, updates: List[dict]) -> None:
|
||||
"""Apply multiple outcome updates in a single read + atomic write.
|
||||
|
||||
Each element of updates must have keys: ticker, trade_date,
|
||||
raw_return, alpha_return, holding_days, reflection.
|
||||
"""
|
||||
if not self._log_path or not self._log_path.exists() or not updates:
|
||||
return
|
||||
|
||||
text = self._log_path.read_text(encoding="utf-8")
|
||||
blocks = text.split(self._SEPARATOR)
|
||||
|
||||
# Build lookup keyed by (trade_date, ticker) for O(1) dispatch
|
||||
update_map = {(u["trade_date"], u["ticker"]): u for u in updates}
|
||||
|
||||
new_blocks = []
|
||||
for block in blocks:
|
||||
stripped = block.strip()
|
||||
if not stripped:
|
||||
new_blocks.append(block)
|
||||
continue
|
||||
|
||||
lines = stripped.splitlines()
|
||||
tag_line = lines[0].strip()
|
||||
|
||||
matched = False
|
||||
for (trade_date, ticker), upd in list(update_map.items()):
|
||||
pending_prefix = f"[{trade_date} | {ticker} |"
|
||||
if tag_line.startswith(pending_prefix) and tag_line.endswith("| pending]"):
|
||||
fields = [f.strip() for f in tag_line[1:-1].split("|")]
|
||||
rating = fields[2]
|
||||
raw_pct = f"{upd['raw_return']:+.1%}"
|
||||
alpha_pct = f"{upd['alpha_return']:+.1%}"
|
||||
new_tag = (
|
||||
f"[{trade_date} | {ticker} | {rating}"
|
||||
f" | {raw_pct} | {alpha_pct} | {upd['holding_days']}d]"
|
||||
)
|
||||
rest = "\n".join(lines[1:])
|
||||
new_blocks.append(
|
||||
f"{new_tag}\n\n{rest.lstrip()}\n\nREFLECTION:\n{upd['reflection']}"
|
||||
)
|
||||
del update_map[(trade_date, ticker)]
|
||||
matched = True
|
||||
break
|
||||
|
||||
if not matched:
|
||||
new_blocks.append(block)
|
||||
|
||||
new_blocks = self._apply_rotation(new_blocks)
|
||||
new_text = self._SEPARATOR.join(new_blocks)
|
||||
tmp_path = self._log_path.with_suffix(".tmp")
|
||||
tmp_path.write_text(new_text, encoding="utf-8")
|
||||
tmp_path.replace(self._log_path)
|
||||
|
||||
# --- Helpers ---
|
||||
|
||||
def _apply_rotation(self, blocks: List[str]) -> List[str]:
|
||||
"""Drop oldest resolved blocks when their count exceeds max_entries.
|
||||
|
||||
Pending blocks are always kept (they represent unprocessed work).
|
||||
Returns ``blocks`` unchanged when rotation is disabled or under cap.
|
||||
"""
|
||||
if not self._max_entries or self._max_entries <= 0:
|
||||
return blocks
|
||||
|
||||
# Tag each block with (kept, is_resolved) by parsing tag-line markers.
|
||||
decisions = []
|
||||
for block in blocks:
|
||||
stripped = block.strip()
|
||||
if not stripped:
|
||||
decisions.append((block, False))
|
||||
continue
|
||||
tag_line = stripped.splitlines()[0].strip()
|
||||
is_resolved = (
|
||||
tag_line.startswith("[")
|
||||
and tag_line.endswith("]")
|
||||
and not tag_line.endswith("| pending]")
|
||||
)
|
||||
decisions.append((block, is_resolved))
|
||||
|
||||
resolved_count = sum(1 for _, r in decisions if r)
|
||||
if resolved_count <= self._max_entries:
|
||||
return blocks
|
||||
|
||||
to_drop = resolved_count - self._max_entries
|
||||
kept: List[str] = []
|
||||
for block, is_resolved in decisions:
|
||||
if is_resolved and to_drop > 0:
|
||||
to_drop -= 1
|
||||
continue
|
||||
kept.append(block)
|
||||
return kept
|
||||
|
||||
def _parse_entry(self, raw: str) -> Optional[dict]:
|
||||
lines = raw.strip().splitlines()
|
||||
if not lines:
|
||||
return None
|
||||
tag_line = lines[0].strip()
|
||||
if not (tag_line.startswith("[") and tag_line.endswith("]")):
|
||||
return None
|
||||
fields = [f.strip() for f in tag_line[1:-1].split("|")]
|
||||
if len(fields) < 4:
|
||||
return None
|
||||
entry = {
|
||||
"date": fields[0],
|
||||
"ticker": fields[1],
|
||||
"rating": fields[2],
|
||||
"pending": fields[3] == "pending",
|
||||
"raw": fields[3] if fields[3] != "pending" else None,
|
||||
"alpha": fields[4] if len(fields) > 4 else None,
|
||||
"holding": fields[5] if len(fields) > 5 else None,
|
||||
}
|
||||
body = "\n".join(lines[1:]).strip()
|
||||
decision_match = self._DECISION_RE.search(body)
|
||||
reflection_match = self._REFLECTION_RE.search(body)
|
||||
entry["decision"] = decision_match.group(1).strip() if decision_match else ""
|
||||
entry["reflection"] = reflection_match.group(1).strip() if reflection_match else ""
|
||||
return entry
|
||||
|
||||
def _format_full(self, e: dict) -> str:
|
||||
raw = e["raw"] or "n/a"
|
||||
alpha = e["alpha"] or "n/a"
|
||||
holding = e["holding"] or "n/a"
|
||||
tag = f"[{e['date']} | {e['ticker']} | {e['rating']} | {raw} | {alpha} | {holding}]"
|
||||
parts = [tag, f"DECISION:\n{e['decision']}"]
|
||||
if e["reflection"]:
|
||||
parts.append(f"REFLECTION:\n{e['reflection']}")
|
||||
return "\n\n".join(parts)
|
||||
|
||||
def _format_reflection_only(self, e: dict) -> str:
|
||||
tag = f"[{e['date']} | {e['ticker']} | {e['rating']} | {e['raw'] or 'n/a'}]"
|
||||
if e["reflection"]:
|
||||
return f"{tag}\n{e['reflection']}"
|
||||
text = e["decision"][:300]
|
||||
suffix = "..." if len(e["decision"]) > 300 else ""
|
||||
return f"{tag}\n{text}{suffix}"
|
||||
|
||||
@@ -0,0 +1,50 @@
|
||||
"""Shared 5-tier rating vocabulary and a deterministic heuristic parser.
|
||||
|
||||
The same five-tier scale (Buy, Overweight, Hold, Underweight, Sell) is used by:
|
||||
- The Research Manager (investment plan recommendation)
|
||||
- The Portfolio Manager (final position decision)
|
||||
- The signal processor (rating extracted for downstream consumers)
|
||||
- The memory log (rating tag stored alongside each decision entry)
|
||||
|
||||
Centralising it here avoids drift between those call sites.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
from typing import Tuple
|
||||
|
||||
|
||||
# Canonical, ordered 5-tier scale (most bullish to most bearish).
|
||||
RATINGS_5_TIER: Tuple[str, ...] = (
|
||||
"Buy", "Overweight", "Hold", "Underweight", "Sell",
|
||||
)
|
||||
|
||||
_RATING_SET = {r.lower() for r in RATINGS_5_TIER}
|
||||
|
||||
# Matches "Rating: X" / "rating - X" / "Rating: **X**" — tolerates markdown
|
||||
# bold wrappers and either a colon or hyphen separator.
|
||||
_RATING_LABEL_RE = re.compile(r"rating.*?[:\-][\s*]*(\w+)", re.IGNORECASE)
|
||||
|
||||
|
||||
def parse_rating(text: str, default: str = "Hold") -> str:
|
||||
"""Heuristically extract a 5-tier rating from prose text.
|
||||
|
||||
Two-pass strategy:
|
||||
1. Look for an explicit "Rating: X" label (tolerant of markdown bold).
|
||||
2. Fall back to the first 5-tier rating word found anywhere in the text.
|
||||
|
||||
Returns a Title-cased rating string, or ``default`` if no rating word appears.
|
||||
"""
|
||||
for line in text.splitlines():
|
||||
m = _RATING_LABEL_RE.search(line)
|
||||
if m and m.group(1).lower() in _RATING_SET:
|
||||
return m.group(1).capitalize()
|
||||
|
||||
for line in text.splitlines():
|
||||
for word in line.lower().split():
|
||||
clean = word.strip("*:.,")
|
||||
if clean in _RATING_SET:
|
||||
return clean.capitalize()
|
||||
|
||||
return default
|
||||
@@ -0,0 +1,73 @@
|
||||
"""Shared helpers for invoking an agent with structured output and a graceful fallback.
|
||||
|
||||
The Portfolio Manager, Trader, and Research Manager all follow the same
|
||||
canonical pattern:
|
||||
|
||||
1. At agent creation, wrap the LLM with ``with_structured_output(Schema)``
|
||||
so the model returns a typed Pydantic instance. If the provider does
|
||||
not support structured output (rare; mostly older Ollama models), the
|
||||
wrap is skipped and the agent uses free-text generation instead.
|
||||
2. At invocation, run the structured call and render the result back to
|
||||
markdown. If the structured call itself fails for any reason
|
||||
(malformed JSON from a weak model, transient provider issue), fall
|
||||
back to a plain ``llm.invoke`` so the pipeline never blocks.
|
||||
|
||||
Centralising the pattern here keeps the agent factories small and ensures
|
||||
all three agents log the same warnings when fallback fires.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
from typing import Any, Callable, Optional, TypeVar
|
||||
|
||||
from pydantic import BaseModel
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
T = TypeVar("T", bound=BaseModel)
|
||||
|
||||
|
||||
def bind_structured(llm: Any, schema: type[T], agent_name: str) -> Optional[Any]:
|
||||
"""Return ``llm.with_structured_output(schema)`` or ``None`` if unsupported.
|
||||
|
||||
Logs a warning when the binding fails so the user understands the agent
|
||||
will use free-text generation for every call instead of one-shot fallback.
|
||||
"""
|
||||
try:
|
||||
return llm.with_structured_output(schema)
|
||||
except (NotImplementedError, AttributeError) as exc:
|
||||
logger.warning(
|
||||
"%s: provider does not support with_structured_output (%s); "
|
||||
"falling back to free-text generation",
|
||||
agent_name, exc,
|
||||
)
|
||||
return None
|
||||
|
||||
|
||||
def invoke_structured_or_freetext(
|
||||
structured_llm: Optional[Any],
|
||||
plain_llm: Any,
|
||||
prompt: Any,
|
||||
render: Callable[[T], str],
|
||||
agent_name: str,
|
||||
) -> str:
|
||||
"""Run the structured call and render to markdown; fall back to free-text on any failure.
|
||||
|
||||
``prompt`` is whatever the underlying LLM accepts (a string for chat
|
||||
invocations, a list of message dicts for chat models that take that
|
||||
shape). The same value is forwarded to the free-text path so the
|
||||
fallback sees the same input the structured call did.
|
||||
"""
|
||||
if structured_llm is not None:
|
||||
try:
|
||||
result = structured_llm.invoke(prompt)
|
||||
return render(result)
|
||||
except Exception as exc:
|
||||
logger.warning(
|
||||
"%s: structured-output invocation failed (%s); retrying once as free text",
|
||||
agent_name, exc,
|
||||
)
|
||||
|
||||
response = plain_llm.invoke(prompt)
|
||||
return response.content
|
||||
@@ -8,6 +8,7 @@ from stockstats import wrap
|
||||
from typing import Annotated
|
||||
import os
|
||||
from .config import get_config
|
||||
from .utils import safe_ticker_component
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
@@ -51,6 +52,10 @@ def load_ohlcv(symbol: str, curr_date: str) -> pd.DataFrame:
|
||||
subsequent calls the cache is reused. Rows after curr_date are
|
||||
filtered out so backtests never see future prices.
|
||||
"""
|
||||
# Reject ticker values that would escape the cache directory when
|
||||
# interpolated into the cache filename (e.g. ``../../tmp/x``).
|
||||
safe_symbol = safe_ticker_component(symbol)
|
||||
|
||||
config = get_config()
|
||||
curr_date_dt = pd.to_datetime(curr_date)
|
||||
|
||||
@@ -63,11 +68,11 @@ def load_ohlcv(symbol: str, curr_date: str) -> pd.DataFrame:
|
||||
os.makedirs(config["data_cache_dir"], exist_ok=True)
|
||||
data_file = os.path.join(
|
||||
config["data_cache_dir"],
|
||||
f"{symbol}-YFin-data-{start_str}-{end_str}.csv",
|
||||
f"{safe_symbol}-YFin-data-{start_str}-{end_str}.csv",
|
||||
)
|
||||
|
||||
if os.path.exists(data_file):
|
||||
data = pd.read_csv(data_file, on_bad_lines="skip")
|
||||
data = pd.read_csv(data_file, on_bad_lines="skip", encoding="utf-8")
|
||||
else:
|
||||
data = yf_retry(lambda: yf.download(
|
||||
symbol,
|
||||
@@ -78,7 +83,7 @@ def load_ohlcv(symbol: str, curr_date: str) -> pd.DataFrame:
|
||||
auto_adjust=True,
|
||||
))
|
||||
data = data.reset_index()
|
||||
data.to_csv(data_file, index=False)
|
||||
data.to_csv(data_file, index=False, encoding="utf-8")
|
||||
|
||||
data = _clean_dataframe(data)
|
||||
|
||||
|
||||
@@ -1,4 +1,5 @@
|
||||
import os
|
||||
import re
|
||||
import json
|
||||
import pandas as pd
|
||||
from datetime import date, timedelta, datetime
|
||||
@@ -6,9 +7,43 @@ from typing import Annotated
|
||||
|
||||
SavePathType = Annotated[str, "File path to save data. If None, data is not saved."]
|
||||
|
||||
# Tickers can contain letters, digits, dot, dash, underscore, and caret
|
||||
# (for index symbols like ^GSPC). Anything else is rejected so the value
|
||||
# never escapes a containing directory when interpolated into a path.
|
||||
_TICKER_PATH_RE = re.compile(r"^[A-Za-z0-9._\-\^]+$")
|
||||
|
||||
|
||||
def safe_ticker_component(value: str, *, max_len: int = 32) -> str:
|
||||
"""Validate ``value`` is safe to interpolate into a filesystem path.
|
||||
|
||||
Tickers come from user CLI input or from LLM tool calls, both of which
|
||||
can be influenced by attacker-controlled content (e.g. prompt injection
|
||||
embedded in fetched news). Without validation, a value like
|
||||
``"../../../etc/foo"`` flows into ``os.path.join`` / ``Path /`` and
|
||||
escapes the configured cache, checkpoint, or results directory.
|
||||
|
||||
Returns ``value`` unchanged when it matches the allowed pattern; raises
|
||||
``ValueError`` otherwise.
|
||||
"""
|
||||
if not isinstance(value, str) or not value:
|
||||
raise ValueError(f"ticker must be a non-empty string, got {value!r}")
|
||||
if len(value) > max_len:
|
||||
raise ValueError(f"ticker exceeds {max_len} chars: {value!r}")
|
||||
if not _TICKER_PATH_RE.fullmatch(value):
|
||||
raise ValueError(
|
||||
f"ticker contains characters not allowed in a filesystem path: {value!r}"
|
||||
)
|
||||
# The regex above allows '.', so values like '.', '..', '...' would pass,
|
||||
# and as a path component they traverse the parent directory. Reject any
|
||||
# value that's only dots.
|
||||
if set(value) == {"."}:
|
||||
raise ValueError(f"ticker cannot consist solely of dots: {value!r}")
|
||||
return value
|
||||
|
||||
|
||||
def save_output(data: pd.DataFrame, tag: str, save_path: SavePathType = None) -> None:
|
||||
if save_path:
|
||||
data.to_csv(save_path)
|
||||
data.to_csv(save_path, encoding="utf-8")
|
||||
print(f"{tag} saved to {save_path}")
|
||||
|
||||
|
||||
|
||||
@@ -6,15 +6,28 @@ DEFAULT_CONFIG = {
|
||||
"project_dir": os.path.abspath(os.path.join(os.path.dirname(__file__), ".")),
|
||||
"results_dir": os.getenv("TRADINGAGENTS_RESULTS_DIR", os.path.join(_TRADINGAGENTS_HOME, "logs")),
|
||||
"data_cache_dir": os.getenv("TRADINGAGENTS_CACHE_DIR", os.path.join(_TRADINGAGENTS_HOME, "cache")),
|
||||
"memory_log_path": os.getenv("TRADINGAGENTS_MEMORY_LOG_PATH", os.path.join(_TRADINGAGENTS_HOME, "memory", "trading_memory.md")),
|
||||
# Optional cap on the number of resolved memory log entries. When set,
|
||||
# the oldest resolved entries are pruned once this limit is exceeded.
|
||||
# Pending entries are never pruned. None disables rotation entirely.
|
||||
"memory_log_max_entries": None,
|
||||
# LLM settings
|
||||
"llm_provider": "openai",
|
||||
"deep_think_llm": "gpt-5.4",
|
||||
"quick_think_llm": "gpt-5.4-mini",
|
||||
"backend_url": "https://api.openai.com/v1",
|
||||
# When None, each provider's client falls back to its own default endpoint
|
||||
# (api.openai.com for OpenAI, generativelanguage.googleapis.com for Gemini, ...).
|
||||
# The CLI overrides this per provider when the user picks one. Keeping a
|
||||
# provider-specific URL here would leak (e.g. OpenAI's /v1 was previously
|
||||
# being forwarded to Gemini, producing malformed request URLs).
|
||||
"backend_url": None,
|
||||
# Provider-specific thinking configuration
|
||||
"google_thinking_level": None, # "high", "minimal", etc.
|
||||
"openai_reasoning_effort": None, # "medium", "high", "low"
|
||||
"anthropic_effort": None, # "high", "medium", "low"
|
||||
# Checkpoint/resume: when True, LangGraph saves state after each node
|
||||
# so a crashed run can resume from the last successful step.
|
||||
"checkpoint_enabled": False,
|
||||
# Output language for analyst reports and final decision
|
||||
# Internal agent debate stays in English for reasoning quality
|
||||
"output_language": "English",
|
||||
|
||||
@@ -0,0 +1,90 @@
|
||||
"""LangGraph checkpoint support for resumable analysis runs.
|
||||
|
||||
Per-ticker SQLite databases so concurrent tickers don't contend.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import hashlib
|
||||
import sqlite3
|
||||
from contextlib import contextmanager
|
||||
from pathlib import Path
|
||||
from typing import Generator
|
||||
|
||||
from langgraph.checkpoint.sqlite import SqliteSaver
|
||||
|
||||
from tradingagents.dataflows.utils import safe_ticker_component
|
||||
|
||||
|
||||
def _db_path(data_dir: str | Path, ticker: str) -> Path:
|
||||
"""Return the SQLite checkpoint DB path for a ticker."""
|
||||
# Reject ticker values that would escape the checkpoints directory.
|
||||
safe = safe_ticker_component(ticker).upper()
|
||||
p = Path(data_dir) / "checkpoints"
|
||||
p.mkdir(parents=True, exist_ok=True)
|
||||
return p / f"{safe}.db"
|
||||
|
||||
|
||||
def thread_id(ticker: str, date: str) -> str:
|
||||
"""Deterministic thread ID for a ticker+date pair."""
|
||||
return hashlib.sha256(f"{ticker.upper()}:{date}".encode()).hexdigest()[:16]
|
||||
|
||||
|
||||
@contextmanager
|
||||
def get_checkpointer(data_dir: str | Path, ticker: str) -> Generator[SqliteSaver, None, None]:
|
||||
"""Context manager yielding a SqliteSaver backed by a per-ticker DB."""
|
||||
db = _db_path(data_dir, ticker)
|
||||
conn = sqlite3.connect(str(db), check_same_thread=False)
|
||||
try:
|
||||
saver = SqliteSaver(conn)
|
||||
saver.setup()
|
||||
yield saver
|
||||
finally:
|
||||
conn.close()
|
||||
|
||||
|
||||
def has_checkpoint(data_dir: str | Path, ticker: str, date: str) -> bool:
|
||||
"""Check whether a resumable checkpoint exists for ticker+date."""
|
||||
return checkpoint_step(data_dir, ticker, date) is not None
|
||||
|
||||
|
||||
def checkpoint_step(data_dir: str | Path, ticker: str, date: str) -> int | None:
|
||||
"""Return the step number of the latest checkpoint, or None if none exists."""
|
||||
db = _db_path(data_dir, ticker)
|
||||
if not db.exists():
|
||||
return None
|
||||
tid = thread_id(ticker, date)
|
||||
with get_checkpointer(data_dir, ticker) as saver:
|
||||
config = {"configurable": {"thread_id": tid}}
|
||||
cp = saver.get_tuple(config)
|
||||
if cp is None:
|
||||
return None
|
||||
return cp.metadata.get("step")
|
||||
|
||||
|
||||
def clear_all_checkpoints(data_dir: str | Path) -> int:
|
||||
"""Remove all checkpoint DBs. Returns number of files deleted."""
|
||||
cp_dir = Path(data_dir) / "checkpoints"
|
||||
if not cp_dir.exists():
|
||||
return 0
|
||||
dbs = list(cp_dir.glob("*.db"))
|
||||
for db in dbs:
|
||||
db.unlink()
|
||||
return len(dbs)
|
||||
|
||||
|
||||
def clear_checkpoint(data_dir: str | Path, ticker: str, date: str) -> None:
|
||||
"""Remove checkpoint for a specific ticker+date by deleting the thread's rows."""
|
||||
db = _db_path(data_dir, ticker)
|
||||
if not db.exists():
|
||||
return
|
||||
tid = thread_id(ticker, date)
|
||||
conn = sqlite3.connect(str(db))
|
||||
try:
|
||||
for table in ("writes", "checkpoints"):
|
||||
conn.execute(f"DELETE FROM {table} WHERE thread_id = ?", (tid,))
|
||||
conn.commit()
|
||||
except sqlite3.OperationalError:
|
||||
pass
|
||||
finally:
|
||||
conn.close()
|
||||
@@ -16,7 +16,11 @@ class Propagator:
|
||||
self.max_recur_limit = max_recur_limit
|
||||
|
||||
def create_initial_state(
|
||||
self, company_name: str, trade_date: str, asset_type: str = "stock"
|
||||
self,
|
||||
company_name: str,
|
||||
trade_date: str,
|
||||
asset_type: str = "stock",
|
||||
past_context: str = "",
|
||||
) -> Dict[str, Any]:
|
||||
"""Create the initial state for the agent graph."""
|
||||
return {
|
||||
@@ -24,6 +28,7 @@ class Propagator:
|
||||
"company_of_interest": company_name,
|
||||
"asset_type": asset_type,
|
||||
"trade_date": str(trade_date),
|
||||
"past_context": past_context,
|
||||
"investment_debate_state": InvestDebateState(
|
||||
{
|
||||
"bull_history": "",
|
||||
|
||||
@@ -1,120 +1,53 @@
|
||||
# TradingAgents/graph/reflection.py
|
||||
|
||||
from typing import Any, Dict
|
||||
from typing import Any
|
||||
|
||||
|
||||
class Reflector:
|
||||
"""Handles reflection on decisions and updating memory."""
|
||||
"""Handles reflection on trading decisions."""
|
||||
|
||||
def __init__(self, quick_thinking_llm: Any):
|
||||
"""Initialize the reflector with an LLM."""
|
||||
self.quick_thinking_llm = quick_thinking_llm
|
||||
self.reflection_system_prompt = self._get_reflection_prompt()
|
||||
self.log_reflection_prompt = self._get_log_reflection_prompt()
|
||||
|
||||
def _get_reflection_prompt(self) -> str:
|
||||
"""Get the system prompt for reflection."""
|
||||
return """
|
||||
You are an expert financial analyst tasked with reviewing trading decisions/analysis and providing a comprehensive, step-by-step analysis.
|
||||
Your goal is to deliver detailed insights into investment decisions and highlight opportunities for improvement, adhering strictly to the following guidelines:
|
||||
def _get_log_reflection_prompt(self) -> str:
|
||||
"""Concise prompt for reflect_on_final_decision (Phase B log entries).
|
||||
|
||||
1. Reasoning:
|
||||
- For each trading decision, determine whether it was correct or incorrect. A correct decision results in an increase in returns, while an incorrect decision does the opposite.
|
||||
- Analyze the contributing factors to each success or mistake. Consider:
|
||||
- Market intelligence.
|
||||
- Technical indicators.
|
||||
- Technical signals.
|
||||
- Price movement analysis.
|
||||
- Overall market data analysis
|
||||
- News analysis.
|
||||
- Social media and sentiment analysis.
|
||||
- Fundamental data analysis.
|
||||
- Weight the importance of each factor in the decision-making process.
|
||||
Produces 2-4 sentences of plain prose — compact enough to be re-injected
|
||||
into future agent prompts without bloating the context window.
|
||||
"""
|
||||
return (
|
||||
"You are a trading analyst reviewing your own past decision now that the outcome is known.\n"
|
||||
"Write exactly 2-4 sentences of plain prose (no bullets, no headers, no markdown).\n\n"
|
||||
"Cover in order:\n"
|
||||
"1. Was the directional call correct? (cite the alpha figure)\n"
|
||||
"2. Which part of the investment thesis held or failed?\n"
|
||||
"3. One concrete lesson to apply to the next similar analysis.\n\n"
|
||||
"Be specific and terse. Your output will be stored verbatim in a decision log "
|
||||
"and re-read by future analysts, so every word must earn its place."
|
||||
)
|
||||
|
||||
2. Improvement:
|
||||
- For any incorrect decisions, propose revisions to maximize returns.
|
||||
- Provide a detailed list of corrective actions or improvements, including specific recommendations (e.g., changing a decision from HOLD to BUY on a particular date).
|
||||
|
||||
3. Summary:
|
||||
- Summarize the lessons learned from the successes and mistakes.
|
||||
- Highlight how these lessons can be adapted for future trading scenarios and draw connections between similar situations to apply the knowledge gained.
|
||||
|
||||
4. Query:
|
||||
- Extract key insights from the summary into a concise sentence of no more than 1000 tokens.
|
||||
- Ensure the condensed sentence captures the essence of the lessons and reasoning for easy reference.
|
||||
|
||||
Adhere strictly to these instructions, and ensure your output is detailed, accurate, and actionable. You will also be given objective descriptions of the market from a price movements, technical indicator, news, and sentiment perspective to provide more context for your analysis.
|
||||
"""
|
||||
|
||||
def _extract_current_situation(self, current_state: Dict[str, Any]) -> str:
|
||||
"""Extract the current market situation from the state."""
|
||||
curr_market_report = current_state["market_report"]
|
||||
curr_sentiment_report = current_state["sentiment_report"]
|
||||
curr_news_report = current_state["news_report"]
|
||||
curr_fundamentals_report = current_state["fundamentals_report"]
|
||||
|
||||
return f"{curr_market_report}\n\n{curr_sentiment_report}\n\n{curr_news_report}\n\n{curr_fundamentals_report}"
|
||||
|
||||
def _reflect_on_component(
|
||||
self, component_type: str, report: str, situation: str, returns_losses
|
||||
def reflect_on_final_decision(
|
||||
self,
|
||||
final_decision: str,
|
||||
raw_return: float,
|
||||
alpha_return: float,
|
||||
) -> str:
|
||||
"""Generate reflection for a component."""
|
||||
"""Single reflection call on the final trade decision with outcome context.
|
||||
|
||||
Used by Phase B deferred reflection. The final_trade_decision already
|
||||
synthesises all analyst insights, so no separate market context is needed.
|
||||
"""
|
||||
messages = [
|
||||
("system", self.reflection_system_prompt),
|
||||
("system", self.log_reflection_prompt),
|
||||
(
|
||||
"human",
|
||||
f"Returns: {returns_losses}\n\nAnalysis/Decision: {report}\n\nObjective Market Reports for Reference: {situation}",
|
||||
(
|
||||
f"Raw return: {raw_return:+.1%}\n"
|
||||
f"Alpha vs SPY: {alpha_return:+.1%}\n\n"
|
||||
f"Final Decision:\n{final_decision}"
|
||||
),
|
||||
),
|
||||
]
|
||||
|
||||
result = self.quick_thinking_llm.invoke(messages).content
|
||||
return result
|
||||
|
||||
def reflect_bull_researcher(self, current_state, returns_losses, bull_memory):
|
||||
"""Reflect on bull researcher's analysis and update memory."""
|
||||
situation = self._extract_current_situation(current_state)
|
||||
bull_debate_history = current_state["investment_debate_state"]["bull_history"]
|
||||
|
||||
result = self._reflect_on_component(
|
||||
"BULL", bull_debate_history, situation, returns_losses
|
||||
)
|
||||
bull_memory.add_situations([(situation, result)])
|
||||
|
||||
def reflect_bear_researcher(self, current_state, returns_losses, bear_memory):
|
||||
"""Reflect on bear researcher's analysis and update memory."""
|
||||
situation = self._extract_current_situation(current_state)
|
||||
bear_debate_history = current_state["investment_debate_state"]["bear_history"]
|
||||
|
||||
result = self._reflect_on_component(
|
||||
"BEAR", bear_debate_history, situation, returns_losses
|
||||
)
|
||||
bear_memory.add_situations([(situation, result)])
|
||||
|
||||
def reflect_trader(self, current_state, returns_losses, trader_memory):
|
||||
"""Reflect on trader's decision and update memory."""
|
||||
situation = self._extract_current_situation(current_state)
|
||||
trader_decision = current_state["trader_investment_plan"]
|
||||
|
||||
result = self._reflect_on_component(
|
||||
"TRADER", trader_decision, situation, returns_losses
|
||||
)
|
||||
trader_memory.add_situations([(situation, result)])
|
||||
|
||||
def reflect_invest_judge(self, current_state, returns_losses, invest_judge_memory):
|
||||
"""Reflect on investment judge's decision and update memory."""
|
||||
situation = self._extract_current_situation(current_state)
|
||||
judge_decision = current_state["investment_debate_state"]["judge_decision"]
|
||||
|
||||
result = self._reflect_on_component(
|
||||
"INVEST JUDGE", judge_decision, situation, returns_losses
|
||||
)
|
||||
invest_judge_memory.add_situations([(situation, result)])
|
||||
|
||||
def reflect_portfolio_manager(self, current_state, returns_losses, portfolio_manager_memory):
|
||||
"""Reflect on portfolio manager's decision and update memory."""
|
||||
situation = self._extract_current_situation(current_state)
|
||||
judge_decision = current_state["risk_debate_state"]["judge_decision"]
|
||||
|
||||
result = self._reflect_on_component(
|
||||
"PORTFOLIO MANAGER", judge_decision, situation, returns_losses
|
||||
)
|
||||
portfolio_manager_memory.add_situations([(situation, result)])
|
||||
return self.quick_thinking_llm.invoke(messages).content
|
||||
|
||||
@@ -18,22 +18,12 @@ class GraphSetup:
|
||||
quick_thinking_llm: Any,
|
||||
deep_thinking_llm: Any,
|
||||
tool_nodes: Dict[str, ToolNode],
|
||||
bull_memory,
|
||||
bear_memory,
|
||||
trader_memory,
|
||||
invest_judge_memory,
|
||||
portfolio_manager_memory,
|
||||
conditional_logic: ConditionalLogic,
|
||||
):
|
||||
"""Initialize with required components."""
|
||||
self.quick_thinking_llm = quick_thinking_llm
|
||||
self.deep_thinking_llm = deep_thinking_llm
|
||||
self.tool_nodes = tool_nodes
|
||||
self.bull_memory = bull_memory
|
||||
self.bear_memory = bear_memory
|
||||
self.trader_memory = trader_memory
|
||||
self.invest_judge_memory = invest_judge_memory
|
||||
self.portfolio_manager_memory = portfolio_manager_memory
|
||||
self.conditional_logic = conditional_logic
|
||||
|
||||
def setup_graph(
|
||||
@@ -85,24 +75,16 @@ class GraphSetup:
|
||||
tool_nodes["fundamentals"] = self.tool_nodes["fundamentals"]
|
||||
|
||||
# Create researcher and manager nodes
|
||||
bull_researcher_node = create_bull_researcher(
|
||||
self.quick_thinking_llm, self.bull_memory
|
||||
)
|
||||
bear_researcher_node = create_bear_researcher(
|
||||
self.quick_thinking_llm, self.bear_memory
|
||||
)
|
||||
research_manager_node = create_research_manager(
|
||||
self.deep_thinking_llm, self.invest_judge_memory
|
||||
)
|
||||
trader_node = create_trader(self.quick_thinking_llm, self.trader_memory)
|
||||
bull_researcher_node = create_bull_researcher(self.quick_thinking_llm)
|
||||
bear_researcher_node = create_bear_researcher(self.quick_thinking_llm)
|
||||
research_manager_node = create_research_manager(self.deep_thinking_llm)
|
||||
trader_node = create_trader(self.quick_thinking_llm)
|
||||
|
||||
# Create risk analysis nodes
|
||||
aggressive_analyst = create_aggressive_debator(self.quick_thinking_llm)
|
||||
neutral_analyst = create_neutral_debator(self.quick_thinking_llm)
|
||||
conservative_analyst = create_conservative_debator(self.quick_thinking_llm)
|
||||
portfolio_manager_node = create_portfolio_manager(
|
||||
self.deep_thinking_llm, self.portfolio_manager_memory
|
||||
)
|
||||
portfolio_manager_node = create_portfolio_manager(self.deep_thinking_llm)
|
||||
|
||||
# Create workflow
|
||||
workflow = StateGraph(AgentState)
|
||||
@@ -197,5 +179,4 @@ class GraphSetup:
|
||||
|
||||
workflow.add_edge("Portfolio Manager", END)
|
||||
|
||||
# Compile and return
|
||||
return workflow.compile()
|
||||
return workflow
|
||||
|
||||
@@ -1,33 +1,31 @@
|
||||
# TradingAgents/graph/signal_processing.py
|
||||
"""Extract the 5-tier portfolio rating from the Portfolio Manager's decision.
|
||||
|
||||
The Portfolio Manager produces a typed ``PortfolioDecision`` via structured
|
||||
output and renders it to markdown that always carries a ``**Rating**: X``
|
||||
header (see :func:`tradingagents.agents.schemas.render_pm_decision`). The
|
||||
deterministic heuristic in :mod:`tradingagents.agents.utils.rating` is more
|
||||
than sufficient to extract that rating; no extra LLM call is needed.
|
||||
|
||||
This module exists for backwards compatibility with callers that expect a
|
||||
``SignalProcessor.process_signal(text)`` interface.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any
|
||||
|
||||
from tradingagents.agents.utils.rating import parse_rating
|
||||
|
||||
|
||||
class SignalProcessor:
|
||||
"""Processes trading signals to extract actionable decisions."""
|
||||
"""Read the 5-tier rating out of a Portfolio Manager decision."""
|
||||
|
||||
def __init__(self, quick_thinking_llm: Any):
|
||||
"""Initialize with an LLM for processing."""
|
||||
def __init__(self, quick_thinking_llm: Any = None):
|
||||
# The LLM argument is accepted for backwards compatibility but no
|
||||
# longer used: the PM's structured output guarantees the rating is
|
||||
# parseable from the rendered markdown without a second LLM call.
|
||||
self.quick_thinking_llm = quick_thinking_llm
|
||||
|
||||
def process_signal(self, full_signal: str) -> str:
|
||||
"""
|
||||
Process a full trading signal to extract the core decision.
|
||||
|
||||
Args:
|
||||
full_signal: Complete trading signal text
|
||||
|
||||
Returns:
|
||||
Extracted rating (BUY, OVERWEIGHT, HOLD, UNDERWEIGHT, or SELL)
|
||||
"""
|
||||
messages = [
|
||||
(
|
||||
"system",
|
||||
"You are an efficient assistant that extracts the trading decision from analyst reports. "
|
||||
"Extract the rating as exactly one of: BUY, OVERWEIGHT, HOLD, UNDERWEIGHT, SELL. "
|
||||
"Output only the single rating word, nothing else.",
|
||||
),
|
||||
("human", full_signal),
|
||||
]
|
||||
|
||||
return self.quick_thinking_llm.invoke(messages).content
|
||||
"""Return one of Buy / Overweight / Hold / Underweight / Sell."""
|
||||
return parse_rating(full_signal)
|
||||
|
||||
@@ -1,18 +1,24 @@
|
||||
# TradingAgents/graph/trading_graph.py
|
||||
|
||||
import logging
|
||||
import os
|
||||
from pathlib import Path
|
||||
import json
|
||||
from datetime import date
|
||||
from datetime import datetime, timedelta
|
||||
from typing import Dict, Any, Tuple, List, Optional
|
||||
|
||||
import yfinance as yf
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
from langgraph.prebuilt import ToolNode
|
||||
|
||||
from tradingagents.llm_clients import create_llm_client
|
||||
|
||||
from tradingagents.agents import *
|
||||
from tradingagents.default_config import DEFAULT_CONFIG
|
||||
from tradingagents.agents.utils.memory import FinancialSituationMemory
|
||||
from tradingagents.agents.utils.memory import TradingMemoryLog
|
||||
from tradingagents.dataflows.utils import safe_ticker_component
|
||||
from tradingagents.agents.utils.agent_states import (
|
||||
AgentState,
|
||||
InvestDebateState,
|
||||
@@ -33,6 +39,7 @@ from tradingagents.agents.utils.agent_utils import (
|
||||
get_global_news
|
||||
)
|
||||
|
||||
from .checkpointer import checkpoint_step, clear_checkpoint, get_checkpointer, thread_id
|
||||
from .conditional_logic import ConditionalLogic
|
||||
from .setup import GraphSetup
|
||||
from .propagation import Propagator
|
||||
@@ -92,12 +99,7 @@ class TradingAgentsGraph:
|
||||
self.deep_thinking_llm = deep_client.get_llm()
|
||||
self.quick_thinking_llm = quick_client.get_llm()
|
||||
|
||||
# Initialize memories
|
||||
self.bull_memory = FinancialSituationMemory("bull_memory", self.config)
|
||||
self.bear_memory = FinancialSituationMemory("bear_memory", self.config)
|
||||
self.trader_memory = FinancialSituationMemory("trader_memory", self.config)
|
||||
self.invest_judge_memory = FinancialSituationMemory("invest_judge_memory", self.config)
|
||||
self.portfolio_manager_memory = FinancialSituationMemory("portfolio_manager_memory", self.config)
|
||||
self.memory_log = TradingMemoryLog(self.config)
|
||||
|
||||
# Create tool nodes
|
||||
self.tool_nodes = self._create_tool_nodes()
|
||||
@@ -111,11 +113,6 @@ class TradingAgentsGraph:
|
||||
self.quick_thinking_llm,
|
||||
self.deep_thinking_llm,
|
||||
self.tool_nodes,
|
||||
self.bull_memory,
|
||||
self.bear_memory,
|
||||
self.trader_memory,
|
||||
self.invest_judge_memory,
|
||||
self.portfolio_manager_memory,
|
||||
self.conditional_logic,
|
||||
)
|
||||
|
||||
@@ -128,8 +125,10 @@ class TradingAgentsGraph:
|
||||
self.ticker = None
|
||||
self.log_states_dict = {} # date to full state dict
|
||||
|
||||
# Set up the graph
|
||||
self.graph = self.graph_setup.setup_graph(selected_analysts)
|
||||
# Set up the graph: keep the workflow for recompilation with a checkpointer.
|
||||
self.workflow = self.graph_setup.setup_graph(selected_analysts)
|
||||
self.graph = self.workflow.compile()
|
||||
self._checkpointer_ctx = None
|
||||
|
||||
def _get_provider_kwargs(self) -> Dict[str, Any]:
|
||||
"""Get provider-specific kwargs for LLM client creation."""
|
||||
@@ -189,19 +188,133 @@ class TradingAgentsGraph:
|
||||
),
|
||||
}
|
||||
|
||||
def propagate(self, company_name, trade_date):
|
||||
"""Run the trading agents graph for a company on a specific date."""
|
||||
def _fetch_returns(
|
||||
self, ticker: str, trade_date: str, holding_days: int = 5
|
||||
) -> Tuple[Optional[float], Optional[float], Optional[int]]:
|
||||
"""Fetch raw and alpha return for ticker over holding_days from trade_date.
|
||||
|
||||
Returns (raw_return, alpha_return, actual_holding_days) or
|
||||
(None, None, None) if price data is unavailable (too recent, delisted,
|
||||
or network error).
|
||||
"""
|
||||
try:
|
||||
start = datetime.strptime(trade_date, "%Y-%m-%d")
|
||||
end = start + timedelta(days=holding_days + 7) # buffer for weekends/holidays
|
||||
end_str = end.strftime("%Y-%m-%d")
|
||||
|
||||
stock = yf.Ticker(ticker).history(start=trade_date, end=end_str)
|
||||
spy = yf.Ticker("SPY").history(start=trade_date, end=end_str)
|
||||
|
||||
if len(stock) < 2 or len(spy) < 2:
|
||||
return None, None, None
|
||||
|
||||
actual_days = min(holding_days, len(stock) - 1, len(spy) - 1)
|
||||
raw = float(
|
||||
(stock["Close"].iloc[actual_days] - stock["Close"].iloc[0])
|
||||
/ stock["Close"].iloc[0]
|
||||
)
|
||||
spy_ret = float(
|
||||
(spy["Close"].iloc[actual_days] - spy["Close"].iloc[0])
|
||||
/ spy["Close"].iloc[0]
|
||||
)
|
||||
alpha = raw - spy_ret
|
||||
return raw, alpha, actual_days
|
||||
except Exception as e:
|
||||
logger.warning(
|
||||
"Could not resolve outcome for %s on %s (will retry next run): %s",
|
||||
ticker, trade_date, e,
|
||||
)
|
||||
return None, None, None
|
||||
|
||||
def _resolve_pending_entries(self, ticker: str) -> None:
|
||||
"""Resolve pending log entries for ticker at the start of a new run.
|
||||
|
||||
Fetches returns for each same-ticker pending entry, generates reflections,
|
||||
then writes all updates in a single atomic batch write to avoid redundant I/O.
|
||||
Skips entries whose price data is not yet available (too recent or delisted).
|
||||
|
||||
Trade-off: only same-ticker entries are resolved per run. Entries for
|
||||
other tickers accumulate until that ticker is run again.
|
||||
"""
|
||||
pending = [e for e in self.memory_log.get_pending_entries() if e["ticker"] == ticker]
|
||||
if not pending:
|
||||
return
|
||||
|
||||
updates = []
|
||||
for entry in pending:
|
||||
raw, alpha, days = self._fetch_returns(ticker, entry["date"])
|
||||
if raw is None:
|
||||
continue # price not available yet — try again next run
|
||||
reflection = self.reflector.reflect_on_final_decision(
|
||||
final_decision=entry.get("decision", ""),
|
||||
raw_return=raw,
|
||||
alpha_return=alpha,
|
||||
)
|
||||
updates.append({
|
||||
"ticker": ticker,
|
||||
"trade_date": entry["date"],
|
||||
"raw_return": raw,
|
||||
"alpha_return": alpha,
|
||||
"holding_days": days,
|
||||
"reflection": reflection,
|
||||
})
|
||||
|
||||
if updates:
|
||||
self.memory_log.batch_update_with_outcomes(updates)
|
||||
|
||||
def propagate(self, company_name, trade_date):
|
||||
"""Run the trading agents graph for a company on a specific date.
|
||||
|
||||
When ``checkpoint_enabled`` is set in config, the graph is recompiled
|
||||
with a per-ticker SqliteSaver so a crashed run can resume from the last
|
||||
successful node on a subsequent invocation with the same ticker+date.
|
||||
"""
|
||||
self.ticker = company_name
|
||||
|
||||
# Initialize state
|
||||
# Resolve any pending memory-log entries for this ticker before the pipeline runs.
|
||||
self._resolve_pending_entries(company_name)
|
||||
|
||||
# Recompile with a checkpointer if the user opted in.
|
||||
if self.config.get("checkpoint_enabled"):
|
||||
self._checkpointer_ctx = get_checkpointer(
|
||||
self.config["data_cache_dir"], company_name
|
||||
)
|
||||
saver = self._checkpointer_ctx.__enter__()
|
||||
self.graph = self.workflow.compile(checkpointer=saver)
|
||||
|
||||
step = checkpoint_step(
|
||||
self.config["data_cache_dir"], company_name, str(trade_date)
|
||||
)
|
||||
if step is not None:
|
||||
logger.info(
|
||||
"Resuming from step %d for %s on %s", step, company_name, trade_date
|
||||
)
|
||||
else:
|
||||
logger.info("Starting fresh for %s on %s", company_name, trade_date)
|
||||
|
||||
try:
|
||||
return self._run_graph(company_name, trade_date)
|
||||
finally:
|
||||
if self._checkpointer_ctx is not None:
|
||||
self._checkpointer_ctx.__exit__(None, None, None)
|
||||
self._checkpointer_ctx = None
|
||||
self.graph = self.workflow.compile()
|
||||
|
||||
def _run_graph(self, company_name, trade_date):
|
||||
"""Execute the graph and write the resulting state to disk and memory log."""
|
||||
# Initialize state — inject memory log context for PM.
|
||||
past_context = self.memory_log.get_past_context(company_name)
|
||||
init_agent_state = self.propagator.create_initial_state(
|
||||
company_name, trade_date
|
||||
company_name, trade_date, past_context=past_context
|
||||
)
|
||||
args = self.propagator.get_graph_args()
|
||||
|
||||
# Inject thread_id so same ticker+date resumes, different date starts fresh.
|
||||
if self.config.get("checkpoint_enabled"):
|
||||
tid = thread_id(company_name, str(trade_date))
|
||||
args.setdefault("config", {}).setdefault("configurable", {})["thread_id"] = tid
|
||||
|
||||
if self.debug:
|
||||
# Debug mode with tracing
|
||||
trace = []
|
||||
for chunk in self.graph.stream(init_agent_state, **args):
|
||||
if len(chunk["messages"]) == 0:
|
||||
@@ -209,19 +322,29 @@ class TradingAgentsGraph:
|
||||
else:
|
||||
chunk["messages"][-1].pretty_print()
|
||||
trace.append(chunk)
|
||||
|
||||
final_state = trace[-1]
|
||||
else:
|
||||
# Standard mode without tracing
|
||||
final_state = self.graph.invoke(init_agent_state, **args)
|
||||
|
||||
# Store current state for reflection
|
||||
# Store current state for reflection.
|
||||
self.curr_state = final_state
|
||||
|
||||
# Log state
|
||||
# Log state to disk.
|
||||
self._log_state(trade_date, final_state)
|
||||
|
||||
# Return decision and processed signal
|
||||
# Store decision for deferred reflection on the next same-ticker run.
|
||||
self.memory_log.store_decision(
|
||||
ticker=company_name,
|
||||
trade_date=trade_date,
|
||||
final_trade_decision=final_state["final_trade_decision"],
|
||||
)
|
||||
|
||||
# Clear checkpoint on successful completion to avoid stale state.
|
||||
if self.config.get("checkpoint_enabled"):
|
||||
clear_checkpoint(
|
||||
self.config["data_cache_dir"], company_name, str(trade_date)
|
||||
)
|
||||
|
||||
return final_state, self.process_signal(final_state["final_trade_decision"])
|
||||
|
||||
def _log_state(self, trade_date, final_state):
|
||||
@@ -256,32 +379,16 @@ class TradingAgentsGraph:
|
||||
"final_trade_decision": final_state["final_trade_decision"],
|
||||
}
|
||||
|
||||
# Save to file
|
||||
directory = Path(self.config["results_dir"]) / self.ticker / "TradingAgentsStrategy_logs"
|
||||
# Save to file. Reject ticker values that would escape the
|
||||
# results directory when joined as a path component.
|
||||
safe_ticker = safe_ticker_component(self.ticker)
|
||||
directory = Path(self.config["results_dir"]) / safe_ticker / "TradingAgentsStrategy_logs"
|
||||
directory.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
log_path = directory / f"full_states_log_{trade_date}.json"
|
||||
with open(log_path, "w", encoding="utf-8") as f:
|
||||
json.dump(self.log_states_dict[str(trade_date)], f, indent=4)
|
||||
|
||||
def reflect_and_remember(self, returns_losses):
|
||||
"""Reflect on decisions and update memory based on returns."""
|
||||
self.reflector.reflect_bull_researcher(
|
||||
self.curr_state, returns_losses, self.bull_memory
|
||||
)
|
||||
self.reflector.reflect_bear_researcher(
|
||||
self.curr_state, returns_losses, self.bear_memory
|
||||
)
|
||||
self.reflector.reflect_trader(
|
||||
self.curr_state, returns_losses, self.trader_memory
|
||||
)
|
||||
self.reflector.reflect_invest_judge(
|
||||
self.curr_state, returns_losses, self.invest_judge_memory
|
||||
)
|
||||
self.reflector.reflect_portfolio_manager(
|
||||
self.curr_state, returns_losses, self.portfolio_manager_memory
|
||||
)
|
||||
|
||||
def process_signal(self, full_signal):
|
||||
"""Process a signal to extract the core decision."""
|
||||
return self.signal_processor.process_signal(full_signal)
|
||||
|
||||
@@ -1,10 +1,6 @@
|
||||
from typing import Optional
|
||||
|
||||
from .base_client import BaseLLMClient
|
||||
from .openai_client import OpenAIClient
|
||||
from .anthropic_client import AnthropicClient
|
||||
from .google_client import GoogleClient
|
||||
from .azure_client import AzureOpenAIClient
|
||||
|
||||
# Providers that use the OpenAI-compatible chat completions API
|
||||
_OPENAI_COMPATIBLE = (
|
||||
@@ -20,6 +16,10 @@ def create_llm_client(
|
||||
) -> BaseLLMClient:
|
||||
"""Create an LLM client for the specified provider.
|
||||
|
||||
Provider modules are imported lazily so that simply importing this
|
||||
factory (e.g. during test collection) does not pull in heavy LLM SDKs
|
||||
or fail when their API keys are absent.
|
||||
|
||||
Args:
|
||||
provider: LLM provider name
|
||||
model: Model name/identifier
|
||||
@@ -35,15 +35,19 @@ def create_llm_client(
|
||||
provider_lower = provider.lower()
|
||||
|
||||
if provider_lower in _OPENAI_COMPATIBLE:
|
||||
from .openai_client import OpenAIClient
|
||||
return OpenAIClient(model, base_url, provider=provider_lower, **kwargs)
|
||||
|
||||
if provider_lower == "anthropic":
|
||||
from .anthropic_client import AnthropicClient
|
||||
return AnthropicClient(model, base_url, **kwargs)
|
||||
|
||||
if provider_lower == "google":
|
||||
from .google_client import GoogleClient
|
||||
return GoogleClient(model, base_url, **kwargs)
|
||||
|
||||
if provider_lower == "azure":
|
||||
from .azure_client import AzureOpenAIClient
|
||||
return AzureOpenAIClient(model, base_url, **kwargs)
|
||||
|
||||
raise ValueError(f"Unsupported LLM provider: {provider}")
|
||||
|
||||
@@ -65,10 +65,12 @@ MODEL_OPTIONS: ProviderModeOptions = {
|
||||
},
|
||||
"deepseek": {
|
||||
"quick": [
|
||||
("DeepSeek V4 Flash - Latest V4 fast model", "deepseek-v4-flash"),
|
||||
("DeepSeek V3.2", "deepseek-chat"),
|
||||
("Custom model ID", "custom"),
|
||||
],
|
||||
"deep": [
|
||||
("DeepSeek V4 Pro - Latest V4 flagship model", "deepseek-v4-pro"),
|
||||
("DeepSeek V3.2 (thinking)", "deepseek-reasoner"),
|
||||
("DeepSeek V3.2", "deepseek-chat"),
|
||||
("Custom model ID", "custom"),
|
||||
|
||||
@@ -1,6 +1,7 @@
|
||||
import os
|
||||
from typing import Any, Optional
|
||||
|
||||
from langchain_core.messages import AIMessage
|
||||
from langchain_openai import ChatOpenAI
|
||||
|
||||
from .base_client import BaseLLMClient, normalize_content
|
||||
@@ -11,13 +12,97 @@ class NormalizedChatOpenAI(ChatOpenAI):
|
||||
"""ChatOpenAI with normalized content output.
|
||||
|
||||
The Responses API returns content as a list of typed blocks
|
||||
(reasoning, text, etc.). This normalizes to string for consistent
|
||||
downstream handling.
|
||||
(reasoning, text, etc.). ``invoke`` normalizes to string for
|
||||
consistent downstream handling. ``with_structured_output`` defaults
|
||||
to function-calling so the Responses-API parse path is avoided
|
||||
(langchain-openai's parse path emits noisy
|
||||
PydanticSerializationUnexpectedValue warnings per call without
|
||||
affecting correctness).
|
||||
|
||||
Provider-specific quirks (e.g. DeepSeek's thinking mode) live in
|
||||
purpose-built subclasses below so this base class stays small.
|
||||
"""
|
||||
|
||||
def invoke(self, input, config=None, **kwargs):
|
||||
return normalize_content(super().invoke(input, config, **kwargs))
|
||||
|
||||
def with_structured_output(self, schema, *, method=None, **kwargs):
|
||||
if method is None:
|
||||
method = "function_calling"
|
||||
return super().with_structured_output(schema, method=method, **kwargs)
|
||||
|
||||
|
||||
def _input_to_messages(input_: Any) -> list:
|
||||
"""Normalise a langchain LLM input to a list of message objects.
|
||||
|
||||
Accepts a list of messages, a ``ChatPromptValue`` (from a
|
||||
ChatPromptTemplate), or anything else (treated as no messages).
|
||||
Used by providers that need to walk the outgoing message history;
|
||||
in particular DeepSeek thinking-mode propagation must work for
|
||||
both bare-list invocations and ChatPromptTemplate-driven ones, so
|
||||
treating only ``list`` here would silently skip half the call sites.
|
||||
"""
|
||||
if isinstance(input_, list):
|
||||
return input_
|
||||
if hasattr(input_, "to_messages"):
|
||||
return input_.to_messages()
|
||||
return []
|
||||
|
||||
|
||||
class DeepSeekChatOpenAI(NormalizedChatOpenAI):
|
||||
"""DeepSeek-specific overrides on top of the OpenAI-compatible client.
|
||||
|
||||
Two quirks that don't apply to other OpenAI-compatible providers:
|
||||
|
||||
1. **Thinking-mode round-trip.** When DeepSeek's thinking models return
|
||||
a response with ``reasoning_content``, that field must be echoed
|
||||
back as part of the assistant message on the next turn or the API
|
||||
fails with HTTP 400. ``_create_chat_result`` captures the field on
|
||||
receive and ``_get_request_payload`` re-attaches it on send.
|
||||
|
||||
2. **deepseek-reasoner has no tool_choice.** Structured output via
|
||||
function-calling is unavailable, so we raise NotImplementedError
|
||||
and let the agent factories fall back to free-text generation
|
||||
(see ``tradingagents/agents/utils/structured.py``).
|
||||
"""
|
||||
|
||||
def _get_request_payload(self, input_, *, stop=None, **kwargs):
|
||||
payload = super()._get_request_payload(input_, stop=stop, **kwargs)
|
||||
outgoing = payload.get("messages", [])
|
||||
for message_dict, message in zip(outgoing, _input_to_messages(input_)):
|
||||
if not isinstance(message, AIMessage):
|
||||
continue
|
||||
reasoning = message.additional_kwargs.get("reasoning_content")
|
||||
if reasoning is not None:
|
||||
message_dict["reasoning_content"] = reasoning
|
||||
return payload
|
||||
|
||||
def _create_chat_result(self, response, generation_info=None):
|
||||
chat_result = super()._create_chat_result(response, generation_info)
|
||||
response_dict = (
|
||||
response
|
||||
if isinstance(response, dict)
|
||||
else response.model_dump(
|
||||
exclude={"choices": {"__all__": {"message": {"parsed"}}}}
|
||||
)
|
||||
)
|
||||
for generation, choice in zip(
|
||||
chat_result.generations, response_dict.get("choices", [])
|
||||
):
|
||||
reasoning = choice.get("message", {}).get("reasoning_content")
|
||||
if reasoning is not None:
|
||||
generation.message.additional_kwargs["reasoning_content"] = reasoning
|
||||
return chat_result
|
||||
|
||||
def with_structured_output(self, schema, *, method=None, **kwargs):
|
||||
if self.model_name == "deepseek-reasoner":
|
||||
raise NotImplementedError(
|
||||
"deepseek-reasoner does not support tool_choice; structured "
|
||||
"output is unavailable. Agent factories fall back to "
|
||||
"free-text generation automatically."
|
||||
)
|
||||
return super().with_structured_output(schema, method=method, **kwargs)
|
||||
|
||||
# Kwargs forwarded from user config to ChatOpenAI
|
||||
_PASSTHROUGH_KWARGS = (
|
||||
"timeout", "max_retries", "reasoning_effort",
|
||||
@@ -59,10 +144,12 @@ class OpenAIClient(BaseLLMClient):
|
||||
self.warn_if_unknown_model()
|
||||
llm_kwargs = {"model": self.model}
|
||||
|
||||
# Provider-specific base URL and auth
|
||||
# Provider-specific base URL and auth. An explicit base_url on the
|
||||
# client (e.g. a corporate proxy) takes precedence over the
|
||||
# provider default so users can route through their own gateway.
|
||||
if self.provider in _PROVIDER_CONFIG:
|
||||
base_url, api_key_env = _PROVIDER_CONFIG[self.provider]
|
||||
llm_kwargs["base_url"] = base_url
|
||||
default_base, api_key_env = _PROVIDER_CONFIG[self.provider]
|
||||
llm_kwargs["base_url"] = self.base_url or default_base
|
||||
if api_key_env:
|
||||
api_key = os.environ.get(api_key_env)
|
||||
if api_key:
|
||||
@@ -82,7 +169,10 @@ class OpenAIClient(BaseLLMClient):
|
||||
if self.provider == "openai":
|
||||
llm_kwargs["use_responses_api"] = True
|
||||
|
||||
return NormalizedChatOpenAI(**llm_kwargs)
|
||||
# DeepSeek's thinking-mode quirks live in their own subclass so the
|
||||
# base NormalizedChatOpenAI stays free of provider-specific branches.
|
||||
chat_cls = DeepSeekChatOpenAI if self.provider == "deepseek" else NormalizedChatOpenAI
|
||||
return chat_cls(**llm_kwargs)
|
||||
|
||||
def validate_model(self) -> bool:
|
||||
"""Validate model for the provider."""
|
||||
|
||||
Reference in New Issue
Block a user