Spaces:
Sleeping
Sleeping
File size: 9,852 Bytes
645673f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 | """Multi-agent investment debate: Pitcher -> Skeptic -> Judge.
The debate replaces the single-LLM analyst call with three specialised
agents that produce a more rigorous, hallucination-resistant verdict.
Toggle: set ``USE_DEBATE=true`` in the environment to enable.
Architecture (LangGraph subgraph):
pitcher_node (Gemma) -> bull_case
skeptic_node (Mistral) -> bear_case
judge_node (Nemotron) -> InvestmentVerdict
"""
import os
import warnings
from typing import TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.types import RetryPolicy
from src.core.logger import get_logger
from src.llm import MODEL_CHAIN
logger = get_logger(__name__)
PITCHER_MODEL = os.getenv("DEBATE_PITCHER_MODEL", "google/gemma-3-27b-it:free")
SKEPTIC_MODEL = os.getenv("DEBATE_SKEPTIC_MODEL", "mistralai/mistral-small-3.1-24b-instruct:free")
JUDGE_MODEL = os.getenv("DEBATE_JUDGE_MODEL", MODEL_CHAIN[0])
class DebateState(TypedDict, total=False):
"""Internal state for the debate subgraph."""
ticker: str
company_name: str
financial_data_summary: str
deep_fundamentals: str
sec_context: str
strategy: str
price: float
eps: float
book_value: float
ebitda: float
bull_case: str
bear_case: str
final_verdict: str
def _make_llm(model: str, max_tokens: int = 2048):
from langchain_openai import ChatOpenAI
return ChatOpenAI(
model=model,
api_key=os.getenv("OPENROUTER_API_KEY"),
base_url="https://openrouter.ai/api/v1",
temperature=0,
max_tokens=max_tokens,
)
# ---------------------------------------------------------------------------
# Node 1 — The Pitcher (bullish thesis)
# ---------------------------------------------------------------------------
def pitcher_node(state: DebateState) -> dict:
"""Build the strongest possible investment thesis for the ticker."""
ticker = state.get("ticker", "")
company = state.get("company_name", ticker)
fundamentals = state.get("deep_fundamentals", "")
sec = state.get("sec_context", "")
price = state.get("price", 0)
eps = state.get("eps", 0)
bv = state.get("book_value", 0)
ebitda = state.get("ebitda", 0)
prompt = (
f"You are a bullish stock pitcher. Write the strongest possible "
f"investment thesis for {company} ({ticker}).\n\n"
f"HARD DATA: Price=${price} | EPS={eps} | Book/Share={bv} | EBITDA={ebitda}\n\n"
)
if fundamentals:
prompt += f"FUNDAMENTALS:\n{fundamentals[:3000]}\n\n"
if sec:
prompt += f"SEC FILINGS:\n{sec[:2000]}\n\n"
prompt += (
"Focus on:\n"
"1. Insider activity (buying signals)\n"
"2. The ONE catalyst that could drive the stock higher\n"
"3. Valuation upside (margin of safety math)\n"
"4. Competitive advantages or turnaround signals\n\n"
"Be specific and data-driven. Only cite facts present in the data above."
)
llm = _make_llm(PITCHER_MODEL)
try:
response = llm.invoke(prompt)
bull_case = response.content
except Exception as exc:
logger.warning("Pitcher (%s) failed: %s — trying fallback", PITCHER_MODEL, exc)
fallback = _make_llm(MODEL_CHAIN[-1])
response = fallback.invoke(prompt)
bull_case = response.content
logger.info("Pitcher delivered bull case for %s (%d chars)", ticker, len(bull_case))
return {"bull_case": bull_case}
# ---------------------------------------------------------------------------
# Node 2 — The Skeptic (bearish challenge)
# ---------------------------------------------------------------------------
def skeptic_node(state: DebateState) -> dict:
"""Challenge the bull case with skeptical analysis grounded in data."""
ticker = state.get("ticker", "")
company = state.get("company_name", ticker)
bull_case = state.get("bull_case", "")
fundamentals = state.get("deep_fundamentals", "")
sec = state.get("sec_context", "")
price = state.get("price", 0)
eps = state.get("eps", 0)
bv = state.get("book_value", 0)
ebitda = state.get("ebitda", 0)
prompt = (
f"You are a skeptical risk analyst. Read the BULL CASE below and "
f"tear it apart for {company} ({ticker}).\n\n"
f"HARD DATA: Price=${price} | EPS={eps} | Book/Share={bv} | EBITDA={ebitda}\n\n"
)
if fundamentals:
prompt += f"FUNDAMENTALS:\n{fundamentals[:3000]}\n\n"
if sec:
prompt += f"SEC FILINGS:\n{sec[:2000]}\n\n"
prompt += (
f"BULL CASE TO CHALLENGE:\n{bull_case[:3000]}\n\n"
"Your job:\n"
"1. Use ONLY the provided data — if a claim has no evidence in the data, "
"call it out as FABRICATED\n"
"2. Identify the biggest risk to an investor\n"
"3. Point out any math errors or unsupported assumptions\n"
"4. State what specific evidence would prove the bear case right\n\n"
"Be thorough but concise."
)
llm = _make_llm(SKEPTIC_MODEL)
try:
response = llm.invoke(prompt)
bear_case = response.content
except Exception as exc:
logger.warning("Skeptic (%s) failed: %s — trying fallback", SKEPTIC_MODEL, exc)
fallback = _make_llm(MODEL_CHAIN[-1])
response = fallback.invoke(prompt)
bear_case = response.content
logger.info("Skeptic delivered bear case for %s (%d chars)", ticker, len(bear_case))
return {"bear_case": bear_case}
# ---------------------------------------------------------------------------
# Node 3 — The Judge (final verdict with structured output)
# ---------------------------------------------------------------------------
def judge_node(state: DebateState) -> dict:
"""Synthesise the debate into a structured InvestmentVerdict."""
from src.models.verdict import InvestmentVerdict
from src.llm import get_structured_llm
ticker = state.get("ticker", "")
company = state.get("company_name", ticker)
bull_case = state.get("bull_case", "")
bear_case = state.get("bear_case", "")
sec = state.get("sec_context", "")
price = state.get("price", 0)
eps = state.get("eps", 0)
bv = state.get("book_value", 0)
ebitda = state.get("ebitda", 0)
strategy = state.get("strategy", "GRAHAM CLASSIC")
prompt = (
f"You are the Chief Investment Officer making the final call on "
f"{company} ({ticker}).\n\n"
f"HARD DATA: Price=${price} | EPS={eps} | Book/Share={bv} | EBITDA={ebitda}\n\n"
f"BULL CASE (from the Pitcher):\n{bull_case[:3000]}\n\n"
f"BEAR CASE (from the Skeptic):\n{bear_case[:3000]}\n\n"
)
if sec:
prompt += f"SEC FILINGS:\n{sec[:2000]}\n\n"
prompt += (
"RULES:\n"
"1. If the Skeptic flagged any claims as FABRICATED, you MUST downgrade\n"
"2. Weight data-backed arguments more heavily\n"
"3. Use strict " + strategy + " math for the quantitative base\n"
"4. Your verdict must be one of: STRONG BUY, BUY, WATCH, AVOID\n\n"
"Produce a structured investment memo with:\n"
"- quantitative_base: Price vs valuation math\n"
"- lynch_pitch: The best data-backed catalyst\n"
"- munger_invert: The key risk from the bear case\n"
"- verdict: Your final call\n"
"- bottom_line: One sentence summary"
)
structured_llm = get_structured_llm(max_tokens=4096).with_structured_output(
InvestmentVerdict
)
with warnings.catch_warnings():
warnings.filterwarnings("ignore", message="Pydantic serializer warnings")
result = structured_llm.invoke(prompt)
verdict_text = result.to_report()
logger.info("Judge delivered verdict for %s: %s", ticker, result.verdict)
return {"final_verdict": verdict_text, "_structured_result": result}
# ---------------------------------------------------------------------------
# Compile the debate subgraph
# ---------------------------------------------------------------------------
_debate_retry = RetryPolicy(max_attempts=3, initial_interval=2.0)
_debate_graph = StateGraph(DebateState)
_debate_graph.add_node("pitcher", pitcher_node, retry=_debate_retry)
_debate_graph.add_node("skeptic", skeptic_node, retry=_debate_retry)
_debate_graph.add_node("judge", judge_node, retry=_debate_retry)
_debate_graph.add_edge(START, "pitcher")
_debate_graph.add_edge("pitcher", "skeptic")
_debate_graph.add_edge("skeptic", "judge")
_debate_graph.add_edge("judge", END)
debate_app = _debate_graph.compile()
def run_debate(
ticker: str,
company_name: str,
financial_data_summary: str,
deep_fundamentals: str,
sec_context: str,
strategy: str,
price: float,
eps: float,
book_value: float,
ebitda: float,
) -> dict:
"""Run the full pitcher -> skeptic -> judge debate for a ticker.
Returns a dict with ``final_verdict`` (str), ``bull_case``, ``bear_case``,
and ``_structured_result`` (InvestmentVerdict).
"""
initial_state: DebateState = {
"ticker": ticker,
"company_name": company_name,
"financial_data_summary": financial_data_summary,
"deep_fundamentals": deep_fundamentals,
"sec_context": sec_context,
"strategy": strategy,
"price": price,
"eps": eps,
"book_value": book_value,
"ebitda": ebitda,
}
result = debate_app.invoke(initial_state)
return {
"final_verdict": result.get("final_verdict", ""),
"bull_case": result.get("bull_case", ""),
"bear_case": result.get("bear_case", ""),
"_structured_result": result.get("_structured_result"),
}
def is_debate_enabled() -> bool:
"""Check if multi-agent debate mode is turned on."""
return os.getenv("USE_DEBATE", "").lower() in ("true", "1", "yes")
|