release: bump version to 0.59.0

backend/app/services/location/llm_fallback.py

@@ -7,8 +7,11 @@ import re
 from dataclasses import dataclass
 from typing import Any, Iterable
 
+from sqlalchemy.ext.asyncio import AsyncSession
+
 from app.core.countries import COUNTRY_ENTRIES, normalize_country
 from app.schemas.ai import SituationalAnalysisRequest
+from app.ai_tasks.prompts import get_effective_prompt
 from app.services.ai_client import AIProviderClient
 from app.services.ai_tools.evidence_store import normalize_search_evidence
 from app.services.ai_tools.web_search import WebSearchClient, WebSearchError
@@ -23,6 +26,8 @@ from app.services.location.text import (
 
 VALID_LLM_PRECISIONS = {"precise", "site", "city"}
 DEFAULT_MIN_CONFIDENCE = 0.55
+LOCATION_NORMALIZE_PROMPT_KEY = "location.factcheck.normalize"
+LOCATION_RESOLVE_PROMPT_KEY = "location.factcheck.resolve"
 MODEL_CONFIDENCE_WEIGHT = 0.25
 _geocode_llm_city = build_default_nominatim_geocoder()
 _LLM_LOCATION_NAME_KEYS = (
@@ -876,6 +881,7 @@ async def _repair_location_payload_from_text(
     raw_text: str,
     query: LocationQuery,
     entity_type: str,
+    db: AsyncSession | None = None,
 ) -> dict[str, Any] | None:
     """Second-pass structure repair for models that answer in prose.
 
@@ -884,12 +890,11 @@ async def _repair_location_payload_from_text(
     """
     if not coerce_str(raw_text):
         return None
+    prompt = await get_effective_prompt(db, LOCATION_NORMALIZE_PROMPT_KEY)
     request = SituationalAnalysisRequest(
         title=f"Normalize location factcheck for {entity_type}",
-        objective=(
-            "Convert the supplied location factcheck text into exactly one strict "
-            "JSON object. Extract only facts present in the text or original query."
-        ),
+        objective=prompt.prompt,
+        system_prompt=prompt.system_prompt or None,
         context={
             "entity_type": entity_type,
             "location_query": _query_context(query),
@@ -929,6 +934,7 @@ async def collect_llm_location_fallback_candidate(
     provider_client: AIProviderClient,
     query: LocationQuery,
     entity_type: str,
+    db: AsyncSession | None = None,
     attempted_queries: Iterable[str] = (),
     search_evidence: list[dict[str, Any]] | None = None,
     min_confidence: float = DEFAULT_MIN_CONFIDENCE,
@@ -946,13 +952,11 @@ async def collect_llm_location_fallback_candidate(
             attempted_queries=[attempt],
             failure_reason="LLM location factcheck skipped: no WebSearch evidence.",
         )
+    prompt = await get_effective_prompt(db, LOCATION_RESOLVE_PROMPT_KEY)
     request = SituationalAnalysisRequest(
         title=f"Location factcheck fallback for {entity_type}",
-        objective=(
-            "Return exactly one JSON object for the most likely physical location. "
-            "Use only fact-checkable public knowledge; return null fields rather "
-            "than guessing when evidence is weak."
-        ),
+        objective=prompt.prompt,
+        system_prompt=prompt.system_prompt or None,
         context={
             "entity_type": entity_type,
             "location_query": _query_context(query),
@@ -1001,6 +1005,7 @@ async def collect_llm_location_fallback_candidate(
             raw_text=response.content,
             query=query,
             entity_type=entity_type,
+            db=db,
         )
     if payload is None:
         payload = _payload_from_free_text(response.content, query=query)