feat: AI category matching — pass all categories to GPT for best-fit selection
Previously: heuristic text matching (sub_type in category.name) — failed for semantic cases like 'Iran Ceasefire Talks' → 'Guerre — Moyen-Orient'. Now: all EventCategories are sent in the prompt (name + type + description). The AI picks matched_category by semantic understanding, returns it in JSON. - Validates returned name against known categories (case-insensitive) - Auto-sets event sub_type when AI finds a match and sub_type was empty - Returns matched_category in evaluate response + shown in UI eval message Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
OpenSquared
@@ -41,12 +41,39 @@ def _get_api_key() -> str:
|
||||
return key
|
||||
|
||||
|
||||
def _build_prompt(event: Dict[str, Any], category_defaults: Optional[List[Dict]] = None) -> str:
|
||||
cat_context = ""
|
||||
if category_defaults:
|
||||
top = sorted(category_defaults, key=lambda x: x.get("sensitivity", 0), reverse=True)[:8]
|
||||
lines = [f" • {d['instrument_id']}: sensibilité {d.get('sensitivity',0):.0%}, direction usuelle = {d.get('typical_direction','?')}" for d in top]
|
||||
cat_context = "\nImpacts par défaut de cette catégorie (ajuste si l'événement est atypique) :\n" + "\n".join(lines)
|
||||
def _build_prompt(
|
||||
event: Dict[str, Any],
|
||||
all_cats: List[Dict],
|
||||
preselected_defaults: Optional[List[Dict]] = None,
|
||||
) -> str:
|
||||
# ── Category selection block ──────────────────────────────────────────────
|
||||
if all_cats:
|
||||
cat_lines = "\n".join(
|
||||
f" {i+1}. [{c['type']}] {c['name']}"
|
||||
+ (f" — {c.get('sub_type','')}" if c.get('sub_type') else "")
|
||||
+ (f" : {c.get('description','')[:80]}" if c.get('description') else "")
|
||||
for i, c in enumerate(all_cats)
|
||||
)
|
||||
cat_selection_block = f"""
|
||||
ÉTAPE 1 — CHOIX DE CATÉGORIE :
|
||||
Voici toutes les catégories disponibles. Choisis celle qui correspond le mieux à cet événement (ou "none" si aucune ne convient) :
|
||||
{cat_lines}
|
||||
|
||||
Retourne le nom EXACT de la catégorie choisie dans le champ "matched_category" du JSON.
|
||||
"""
|
||||
else:
|
||||
cat_selection_block = ""
|
||||
|
||||
# ── Default impacts from pre-selected category (fallback heuristic) ───────
|
||||
defaults_block = ""
|
||||
if preselected_defaults:
|
||||
top = sorted(preselected_defaults, key=lambda x: x.get("sensitivity", 0), reverse=True)[:8]
|
||||
lines = [
|
||||
f" • {d['instrument_id']}: sensibilité {d.get('sensitivity',0):.0%}, "
|
||||
f"direction usuelle = {d.get('typical_direction','?')}"
|
||||
for d in top
|
||||
]
|
||||
defaults_block = "\nImpacts par défaut (ajuste si l'événement est atypique) :\n" + "\n".join(lines)
|
||||
|
||||
return f"""Tu es un analyste macro senior spécialisé en options et trading multi-actifs.
|
||||
|
||||
@@ -56,10 +83,11 @@ def _build_prompt(event: Dict[str, Any], category_defaults: Optional[List[Dict]]
|
||||
- Type : {event.get('category', '')} / {event.get('sub_type', '')}
|
||||
- Description : {event.get('description', '')}
|
||||
- Impact attendu (résumé) : {event.get('market_impact', '')}
|
||||
{cat_context}
|
||||
{cat_selection_block}{defaults_block}
|
||||
|
||||
MISSION :
|
||||
Évalue l'impact potentiel de cet événement sur chacun des 20 instruments suivants.
|
||||
ÉTAPE 2 — IMPACTS PAR INSTRUMENT :
|
||||
Évalue l'impact potentiel de cet événement sur chacun des 20 instruments.
|
||||
Si tu as choisi une catégorie à l'étape 1, utilise ses impacts par défaut comme point de départ et ajuste selon le contexte spécifique de l'événement.
|
||||
Instruments disponibles : {_INST_LIST_STR}
|
||||
|
||||
Pour chaque instrument IMPACTÉ (score >= 0.15), retourne :
|
||||
@@ -76,10 +104,11 @@ Critères de scoring :
|
||||
- 0.6-0.8 : fort (1-2% mouvement)
|
||||
- 0.8-1.0 : majeur (> 2%, event driver primaire)
|
||||
|
||||
FORMAT JSON STRICT (pas de texte hors du JSON) :
|
||||
FORMAT JSON STRICT :
|
||||
{{
|
||||
"matched_category": "Nom exact de la catégorie choisie ou null",
|
||||
"impacts": [
|
||||
{{"instrument_id": "SPY", "impact_score": 0.85, "direction": "bearish", "rationale": "Hausse taux comprime les multiples P/E", "confidence": 0.90}}
|
||||
{{"instrument_id": "SPY", "impact_score": 0.85, "direction": "bearish", "rationale": "...", "confidence": 0.90}}
|
||||
],
|
||||
"summary": "Contexte macro de cet événement en 1 phrase",
|
||||
"key_driver": "Instrument le plus impacté et pourquoi"
|
||||
@@ -121,46 +150,33 @@ def evaluate_event_impacts(
|
||||
if force and existing:
|
||||
delete_impacts_for_source("event", event_id)
|
||||
|
||||
# Get category defaults — try exact name match first, then fuzzy by sub_type/type
|
||||
category_defaults = None
|
||||
from services.database import get_all_event_categories
|
||||
all_cats = get_all_event_categories()
|
||||
all_cats_raw = get_all_event_categories()
|
||||
|
||||
def _find_cat(candidates):
|
||||
"""Returns (category_name, default_impacts) or (None, None)."""
|
||||
for c in candidates:
|
||||
if c is None:
|
||||
continue
|
||||
row = next((x for x in all_cats if x["name"].lower() == c.lower()), None)
|
||||
if row:
|
||||
try:
|
||||
return row["name"], json.loads(row.get("default_impacts") or "[]")
|
||||
except Exception:
|
||||
pass
|
||||
return None, None
|
||||
# Parse default_impacts JSON for each category
|
||||
all_cats: List[Dict] = []
|
||||
for c in all_cats_raw:
|
||||
try:
|
||||
c["default_impacts"] = json.loads(c.get("default_impacts") or "[]")
|
||||
except Exception:
|
||||
c["default_impacts"] = []
|
||||
all_cats.append(c)
|
||||
|
||||
sub = (event.get("sub_type") or "").strip()
|
||||
cat = (event.get("category") or "").strip()
|
||||
name = (event.get("name") or "").upper()
|
||||
|
||||
# Priority: exact sub_type name → fuzzy keyword in category name → type match
|
||||
candidates = [sub]
|
||||
for row in all_cats:
|
||||
rname = row["name"].upper()
|
||||
if sub and sub.upper() in rname:
|
||||
candidates.append(row["name"])
|
||||
elif any(k in name for k in ["FOMC", "FED"] if k in rname):
|
||||
candidates.append(row["name"])
|
||||
elif row.get("type") == cat and not sub:
|
||||
candidates.append(row["name"])
|
||||
|
||||
matched_cat_name, category_defaults = _find_cat(candidates)
|
||||
# Heuristic pre-selection: if sub_type matches a category exactly, pass its
|
||||
# defaults as a hint (the AI will still pick the best category independently)
|
||||
sub = (event.get("sub_type") or "").strip()
|
||||
preselected_defaults: Optional[List[Dict]] = None
|
||||
if sub:
|
||||
pre = next((c for c in all_cats if c["name"].lower() == sub.lower()), None)
|
||||
if pre:
|
||||
preselected_defaults = pre["default_impacts"]
|
||||
|
||||
api_key = _get_api_key()
|
||||
if not api_key:
|
||||
return {"error": "No OpenAI API key configured"}
|
||||
|
||||
prompt = _build_prompt(event, category_defaults)
|
||||
prompt = _build_prompt(event, all_cats, preselected_defaults)
|
||||
matched_cat_name: Optional[str] = None
|
||||
|
||||
try:
|
||||
from openai import OpenAI
|
||||
@@ -178,9 +194,32 @@ def evaluate_event_impacts(
|
||||
return {"error": str(e)}
|
||||
|
||||
ai_impacts = raw.get("impacts", [])
|
||||
summary = raw.get("summary", "")
|
||||
summary = raw.get("summary", "")
|
||||
key_driver = raw.get("key_driver", "")
|
||||
|
||||
# ── Use AI-chosen category ────────────────────────────────────────────────
|
||||
ai_cat_name = (raw.get("matched_category") or "").strip()
|
||||
if ai_cat_name and ai_cat_name.lower() != "none":
|
||||
# Validate against known categories (case-insensitive)
|
||||
matched = next((c for c in all_cats if c["name"].lower() == ai_cat_name.lower()), None)
|
||||
if matched:
|
||||
matched_cat_name = matched["name"]
|
||||
logger.info(f"[impact_service] AI matched event {event_id} → category '{matched_cat_name}'")
|
||||
# Auto-update event sub_type if not already set
|
||||
current_sub = (event.get("sub_type") or "").strip()
|
||||
if not current_sub:
|
||||
try:
|
||||
from services.database import update_market_event
|
||||
update_market_event(event_id, {**event, "sub_type": matched_cat_name,
|
||||
"source_refs": event.get("source_refs", [])})
|
||||
logger.info(f"[impact_service] Auto-set sub_type='{matched_cat_name}' on event {event_id}")
|
||||
except Exception as _e:
|
||||
logger.warning(f"[impact_service] Could not auto-set sub_type: {_e}")
|
||||
else:
|
||||
logger.warning(f"[impact_service] AI returned unknown category '{ai_cat_name}' for event {event_id}")
|
||||
|
||||
cat_label = matched_cat_name or (event.get("sub_type") or "").strip() or (event.get("category") or "")
|
||||
|
||||
# Persist
|
||||
to_save = []
|
||||
for imp in ai_impacts:
|
||||
@@ -191,7 +230,7 @@ def evaluate_event_impacts(
|
||||
"source_id": event_id,
|
||||
"source_name": event["name"],
|
||||
"source_date": event["start_date"],
|
||||
"category_name": matched_cat_name or sub or cat,
|
||||
"category_name": cat_label,
|
||||
"instrument_id": imp["instrument_id"],
|
||||
"impact_score": float(imp.get("impact_score", 0.5)),
|
||||
"direction": imp.get("direction", "neutral"),
|
||||
@@ -204,13 +243,14 @@ def evaluate_event_impacts(
|
||||
saved = get_impacts_for_source("event", event_id)
|
||||
|
||||
return {
|
||||
"event_id": event_id,
|
||||
"event_name": event["name"],
|
||||
"impacts": saved,
|
||||
"summary": summary,
|
||||
"key_driver": key_driver,
|
||||
"n_instruments": len(saved),
|
||||
"from_cache": False,
|
||||
"event_id": event_id,
|
||||
"event_name": event["name"],
|
||||
"matched_category": matched_cat_name,
|
||||
"impacts": saved,
|
||||
"summary": summary,
|
||||
"key_driver": key_driver,
|
||||
"n_instruments": len(saved),
|
||||
"from_cache": False,
|
||||
}
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user