Initial commit: CI Agent competitive-intelligence monitoring app

FastAPI + Celery + Next.js + Postgres/Redis app with company monitoring,
source collection, LLM-based change analysis, enrichment, and account
security (Turnstile, escalating lockout, email verification).

apps/api/app/prompts/extraction.py

@@ -0,0 +1,54 @@
+"""Task B: fact and signal extraction. Pulls discrete, evidence-anchored
+signals (hiring, partnerships, leadership, financial, etc.) out of a single
+document - the atomic units Task C (synthesis) and Task D (report
+generation) later combine."""
+
+from __future__ import annotations
+
+from pydantic import BaseModel, Field
+
+from app.analysis.llm.base import LLMProvider
+from app.prompts.base import build_user_prompt
+
+SYSTEM_PROMPT = (
+    "You are a competitive intelligence analyst extracting discrete factual signals from a "
+    "single document. Extract only what the text actually states or clearly implies - never "
+    "add outside knowledge. Every signal must include the exact passage that supports it."
+)
+
+_SIGNAL_TYPES = (
+    "event, entity, date, location, product, leadership, partnership, hiring, investment, "
+    "manufacturing, technology, financial, regulatory, sentiment"
+)
+
+
+class ExtractedSignal(BaseModel):
+    signal_type: str = Field(description=f"One of: {_SIGNAL_TYPES}")
+    description: str
+    supporting_passage: str = Field(description="The exact quote from the document backing this")
+    date: str | None = None
+    entities: list[str] = Field(default_factory=list)
+
+
+class ExtractionResult(BaseModel):
+    signals: list[ExtractedSignal] = Field(default_factory=list)
+
+
+async def extract_signals(
+    llm: LLMProvider,
+    *,
+    document_title: str | None,
+    document_url: str,
+    document_text: str,
+) -> ExtractionResult:
+    evidence = {
+        "document_title": document_title,
+        "document_url": document_url,
+        "document_text": document_text[:6000],
+    }
+    user_prompt = build_user_prompt(
+        "Extract every discrete factual signal from this document, each anchored to its "
+        "exact supporting passage.",
+        evidence,
+    )
+    return await llm.generate_structured(SYSTEM_PROMPT, user_prompt, ExtractionResult)