feat: Kontextabhängige Karten-Kategorien
4 feste Farbstufen (primary/secondary/tertiary/mentioned) mit variablen Labels pro Lage, die von Haiku generiert werden. - DB: category_labels Spalte in incidents, alte Kategorien migriert (target->primary, response/retaliation->secondary, actor->tertiary) - Geoparsing: generate_category_labels() + neuer Prompt mit neuen Keys - QC: Kategorieprüfung auf neue Keys umgestellt - Orchestrator: Tuple-Rückgabe + Labels in DB speichern - API: category_labels im Locations- und Lagebild-Response - Frontend: Dynamische Legende aus API-Labels mit Fallback-Defaults - Migrationsskript für bestehende Lagen Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Dieser Commit ist enthalten in:
Claude Dev
@@ -209,6 +209,90 @@ def _geocode_location(name: str, country_code: str = "", haiku_coords: Optional[
|
||||
return result
|
||||
|
||||
|
||||
# Default-Labels (Fallback wenn Haiku keine generiert)
|
||||
DEFAULT_CATEGORY_LABELS = {
|
||||
"primary": "Hauptgeschehen",
|
||||
"secondary": "Reaktionen",
|
||||
"tertiary": "Beteiligte",
|
||||
"mentioned": "Erwaehnt",
|
||||
}
|
||||
|
||||
CATEGORY_LABELS_PROMPT = """Generiere kurze, praegnante Kategorie-Labels fuer Karten-Pins zu dieser Nachrichtenlage.
|
||||
|
||||
Lage: "{incident_context}"
|
||||
|
||||
Es gibt 4 Farbstufen fuer Orte auf der Karte:
|
||||
1. primary (Rot): Wo das Hauptgeschehen stattfindet
|
||||
2. secondary (Orange): Direkte Reaktionen/Gegenmassnahmen
|
||||
3. tertiary (Blau): Entscheidungstraeger/Beteiligte
|
||||
4. mentioned (Grau): Nur erwaehnt
|
||||
|
||||
Generiere fuer jede Stufe ein kurzes Label (1-3 Woerter), das zum Thema passt.
|
||||
Wenn eine Stufe fuer dieses Thema nicht sinnvoll ist, setze null.
|
||||
|
||||
Beispiele:
|
||||
- Militaerkonflikt Iran: {{"primary": "Kampfschauplätze", "secondary": "Vergeltungsschläge", "tertiary": "Strategische Akteure", "mentioned": "Erwähnt"}}
|
||||
- Erdbeben Tuerkei: {{"primary": "Katastrophenzone", "secondary": "Hilfsoperationen", "tertiary": "Geberländer", "mentioned": "Erwähnt"}}
|
||||
- Bundestagswahl: {{"primary": "Wahlkreise", "secondary": "Koalitionspartner", "tertiary": "Internationale Reaktionen", "mentioned": "Erwähnt"}}
|
||||
|
||||
Antworte NUR als JSON-Objekt:"""
|
||||
|
||||
|
||||
async def generate_category_labels(incident_context: str) -> dict[str, str | None]:
|
||||
"""Generiert kontextabhaengige Kategorie-Labels via Haiku.
|
||||
|
||||
Args:
|
||||
incident_context: Lage-Titel + Beschreibung
|
||||
|
||||
Returns:
|
||||
Dict mit Labels fuer primary/secondary/tertiary/mentioned (oder None wenn nicht passend)
|
||||
"""
|
||||
if not incident_context or not incident_context.strip():
|
||||
return dict(DEFAULT_CATEGORY_LABELS)
|
||||
|
||||
prompt = CATEGORY_LABELS_PROMPT.format(incident_context=incident_context[:500])
|
||||
|
||||
try:
|
||||
result_text, usage = await call_claude(prompt, tools=None, model=CLAUDE_MODEL_FAST)
|
||||
parsed = None
|
||||
try:
|
||||
parsed = json.loads(result_text)
|
||||
except json.JSONDecodeError:
|
||||
match = re.search(r'\{.*\}', result_text, re.DOTALL)
|
||||
if match:
|
||||
try:
|
||||
parsed = json.loads(match.group())
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
|
||||
if not parsed or not isinstance(parsed, dict):
|
||||
logger.warning("generate_category_labels: Kein gueltiges JSON erhalten")
|
||||
return dict(DEFAULT_CATEGORY_LABELS)
|
||||
|
||||
# Validierung: Nur erlaubte Keys, Werte muessen str oder None sein
|
||||
valid_keys = {"primary", "secondary", "tertiary", "mentioned"}
|
||||
labels = {}
|
||||
for key in valid_keys:
|
||||
val = parsed.get(key)
|
||||
if val is None or val == "null":
|
||||
labels[key] = None
|
||||
elif isinstance(val, str) and val.strip():
|
||||
labels[key] = val.strip()
|
||||
else:
|
||||
labels[key] = DEFAULT_CATEGORY_LABELS.get(key)
|
||||
|
||||
# mentioned sollte immer einen Wert haben
|
||||
if not labels.get("mentioned"):
|
||||
labels["mentioned"] = "Erwaehnt"
|
||||
|
||||
logger.info(f"Kategorie-Labels generiert: {labels}")
|
||||
return labels
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"generate_category_labels fehlgeschlagen: {e}")
|
||||
return dict(DEFAULT_CATEGORY_LABELS)
|
||||
|
||||
|
||||
HAIKU_GEOPARSE_PROMPT = """Extrahiere alle geographischen Orte aus diesen Nachrichten-Headlines.
|
||||
|
||||
Kontext der Lage: "{incident_context}"
|
||||
@@ -222,9 +306,9 @@ Regeln:
|
||||
- Regionen wie "Middle East", "Gulf", "Naher Osten" NICHT extrahieren (kein einzelner Punkt auf der Karte)
|
||||
|
||||
Klassifiziere basierend auf dem Lage-Kontext:
|
||||
- "target": Wo das Ereignis passiert / Schaden entsteht
|
||||
- "response": Wo Reaktionen / Gegenmassnahmen stattfinden
|
||||
- "actor": Wo Entscheidungen getroffen werden / Entscheider sitzen
|
||||
- "primary": Wo das Hauptgeschehen stattfindet (z.B. Angriffsziele, Katastrophenzone, Wahlkreise)
|
||||
- "secondary": Direkte Reaktionen oder Gegenmassnahmen (z.B. Vergeltung, Hilfsoperationen)
|
||||
- "tertiary": Entscheidungstraeger, Beteiligte (z.B. wo Entscheidungen getroffen werden)
|
||||
- "mentioned": Nur erwaehnt, kein direkter Bezug
|
||||
|
||||
Headlines:
|
||||
@@ -233,7 +317,7 @@ Headlines:
|
||||
Antwort NUR als JSON-Array, kein anderer Text:
|
||||
[{{"headline_idx": 0, "locations": [
|
||||
{{"name": "Teheran", "normalized": "Tehran", "country_code": "IR",
|
||||
"type": "city", "category": "target",
|
||||
"type": "city", "category": "primary",
|
||||
"lat": 35.69, "lon": 51.42}}
|
||||
]}}]"""
|
||||
|
||||
@@ -314,12 +398,19 @@ async def _extract_locations_haiku(
|
||||
if not name:
|
||||
continue
|
||||
|
||||
raw_cat = loc.get("category", "mentioned")
|
||||
# Alte Kategorien mappen (falls Haiku sie noch generiert)
|
||||
cat_map = {"target": "primary", "response": "secondary", "retaliation": "secondary", "actor": "tertiary", "context": "tertiary"}
|
||||
category = cat_map.get(raw_cat, raw_cat)
|
||||
if category not in ("primary", "secondary", "tertiary", "mentioned"):
|
||||
category = "mentioned"
|
||||
|
||||
article_locs.append({
|
||||
"name": name,
|
||||
"normalized": loc.get("normalized", name),
|
||||
"country_code": loc.get("country_code", ""),
|
||||
"type": loc_type,
|
||||
"category": loc.get("category", "mentioned"),
|
||||
"category": category,
|
||||
"lat": loc.get("lat"),
|
||||
"lon": loc.get("lon"),
|
||||
})
|
||||
@@ -333,7 +424,7 @@ async def _extract_locations_haiku(
|
||||
async def geoparse_articles(
|
||||
articles: list[dict],
|
||||
incident_context: str = "",
|
||||
) -> dict[int, list[dict]]:
|
||||
) -> tuple[dict[int, list[dict]], dict[str, str | None] | None]:
|
||||
"""Geoparsing fuer eine Liste von Artikeln via Haiku + geonamescache.
|
||||
|
||||
Args:
|
||||
@@ -341,11 +432,15 @@ async def geoparse_articles(
|
||||
incident_context: Lage-Kontext (Titel + Beschreibung) fuer kontextbewusste Klassifizierung
|
||||
|
||||
Returns:
|
||||
dict[article_id -> list[{location_name, location_name_normalized, country_code,
|
||||
lat, lon, confidence, source_text, category}]]
|
||||
Tuple von (dict[article_id -> list[locations]], category_labels oder None)
|
||||
"""
|
||||
if not articles:
|
||||
return {}
|
||||
return {}, None
|
||||
|
||||
# Labels parallel zum Geoparsing generieren (nur wenn Kontext vorhanden)
|
||||
labels_task = None
|
||||
if incident_context:
|
||||
labels_task = asyncio.create_task(generate_category_labels(incident_context))
|
||||
|
||||
# Headlines sammeln
|
||||
headlines = []
|
||||
@@ -363,7 +458,13 @@ async def geoparse_articles(
|
||||
headlines.append({"idx": article_id, "text": headline})
|
||||
|
||||
if not headlines:
|
||||
return {}
|
||||
category_labels = None
|
||||
if labels_task:
|
||||
try:
|
||||
category_labels = await labels_task
|
||||
except Exception:
|
||||
pass
|
||||
return {}, category_labels
|
||||
|
||||
# Batches bilden (max 50 Headlines pro Haiku-Call)
|
||||
batch_size = 50
|
||||
@@ -374,7 +475,13 @@ async def geoparse_articles(
|
||||
all_haiku_results.update(batch_results)
|
||||
|
||||
if not all_haiku_results:
|
||||
return {}
|
||||
category_labels = None
|
||||
if labels_task:
|
||||
try:
|
||||
category_labels = await labels_task
|
||||
except Exception:
|
||||
pass
|
||||
return {}, category_labels
|
||||
|
||||
# Geocoding via geonamescache (mit Haiku-Koordinaten als Fallback)
|
||||
result = {}
|
||||
@@ -406,4 +513,12 @@ async def geoparse_articles(
|
||||
if locations:
|
||||
result[article_id] = locations
|
||||
|
||||
return result
|
||||
# Category-Labels abwarten
|
||||
category_labels = None
|
||||
if labels_task:
|
||||
try:
|
||||
category_labels = await labels_task
|
||||
except Exception as e:
|
||||
logger.warning(f"Category-Labels konnten nicht generiert werden: {e}")
|
||||
|
||||
return result, category_labels
|
||||
|
||||
@@ -782,7 +782,7 @@ class AgentOrchestrator:
|
||||
from agents.geoparsing import geoparse_articles
|
||||
incident_context = f"{title} - {description}"
|
||||
logger.info(f"Geoparsing fuer {len(new_articles_for_analysis)} neue Artikel...")
|
||||
geo_results = await geoparse_articles(new_articles_for_analysis, incident_context)
|
||||
geo_results, category_labels = await geoparse_articles(new_articles_for_analysis, incident_context)
|
||||
geo_count = 0
|
||||
for art_id, locations in geo_results.items():
|
||||
for loc in locations:
|
||||
@@ -799,6 +799,15 @@ class AgentOrchestrator:
|
||||
if geo_count > 0:
|
||||
await db.commit()
|
||||
logger.info(f"Geoparsing: {geo_count} Orte aus {len(geo_results)} Artikeln gespeichert")
|
||||
# Category-Labels in Incident speichern (nur wenn neu generiert)
|
||||
if category_labels:
|
||||
import json as _json
|
||||
await db.execute(
|
||||
"UPDATE incidents SET category_labels = ? WHERE id = ? AND category_labels IS NULL",
|
||||
(_json.dumps(category_labels, ensure_ascii=False), incident_id),
|
||||
)
|
||||
await db.commit()
|
||||
logger.info(f"Category-Labels gespeichert fuer Incident {incident_id}: {category_labels}")
|
||||
except Exception as e:
|
||||
logger.warning(f"Geoparsing fehlgeschlagen (Pipeline laeuft weiter): {e}")
|
||||
|
||||
|
||||
In neuem Issue referenzieren
Einen Benutzer sperren