release: bump version to 0.62.0

backend/app/services/location/llm_fallback.py

@@ -9,6 +9,7 @@ from typing import Any, Iterable
 
 from sqlalchemy.ext.asyncio import AsyncSession
 
+from app.core.logging import get_logger
 from app.core.countries import COUNTRY_ENTRIES, normalize_country
 from app.schemas.ai import SituationalAnalysisRequest
 from app.ai_tasks.prompts import get_effective_prompt
@@ -29,6 +30,9 @@ DEFAULT_MIN_CONFIDENCE = 0.55
 LOCATION_NORMALIZE_PROMPT_KEY = "location.factcheck.normalize"
 LOCATION_RESOLVE_PROMPT_KEY = "location.factcheck.resolve"
 MODEL_CONFIDENCE_WEIGHT = 0.25
+LOG_TEXT_LIMIT = 1200
+LOG_EVIDENCE_LIMIT = 5
+logger = get_logger(__name__, service="location")
 _geocode_llm_city = build_default_nominatim_geocoder()
 _LLM_LOCATION_NAME_KEYS = (
     "matched_location_name",
@@ -97,6 +101,35 @@ class LocationEvidenceScore:
     summary: str
 
 
+def _truncate_log_text(value: Any, limit: int = LOG_TEXT_LIMIT) -> str:
+    text = coerce_str(value)
+    if len(text) <= limit:
+        return text
+    return f"{text[:limit]}…"
+
+
+def _summarize_search_evidence(evidence: list[dict[str, Any]] | None) -> list[dict[str, Any]]:
+    items: list[dict[str, Any]] = []
+    for item in (evidence or [])[:LOG_EVIDENCE_LIMIT]:
+        if not isinstance(item, dict):
+            continue
+        items.append(
+            {
+                "title": _truncate_log_text(item.get("title"), 180),
+                "source": _truncate_log_text(item.get("source") or item.get("name"), 120),
+                "url": _truncate_log_text(item.get("url"), 240),
+                "snippet": _truncate_log_text(
+                    item.get("snippet")
+                    or item.get("content")
+                    or item.get("text")
+                    or item.get("summary"),
+                    360,
+                ),
+            }
+        )
+    return items
+
+
 def _first_json_object(text: str) -> dict[str, Any] | None:
     stripped = text.strip()
     if not stripped:
@@ -162,6 +195,64 @@ def _evidence_label(item: Any) -> str:
     return coerce_str(item)
 
 
+def _evidence_text(item: dict[str, Any]) -> str:
+    return " ".join(
+        coerce_str(item.get(key))
+        for key in ("title", "source", "name", "url", "snippet", "content", "text", "quote", "summary")
+        if coerce_str(item.get(key))
+    )
+
+
+def _search_evidence_entity_match(item: dict[str, Any], query: LocationQuery) -> bool:
+    haystack = normalize_text(_evidence_text(item))
+    if not haystack:
+        return False
+    needles = [
+        coerce_str(query.name),
+        *[coerce_str(alias) for alias in query.aliases],
+    ]
+    return any(normalize_text(needle) and normalize_text(needle) in haystack for needle in needles)
+
+
+def _evidence_has_location_assertion(item: dict[str, Any], city: str) -> bool:
+    normalized_city = normalize_text(city)
+    text = normalize_text(_evidence_text(item))
+    if not normalized_city or normalized_city not in text:
+        return False
+    assertion_terms = (
+        "located",
+        "situated",
+        "built",
+        "hosted",
+        "deployed",
+        "installed",
+        "facility",
+        "campus",
+        "site",
+        "data center",
+        "datacenter",
+        "supercomputer center",
+        "位于",
+        "位於",
+        "坐落",
+        "建置",
+        "設置",
+        "设置",
+    )
+    return any(term in text for term in assertion_terms)
+
+
+def _city_is_unsupported_name_hint(payload: dict[str, Any], query: LocationQuery, evidence_items: list[dict[str, Any]]) -> bool:
+    city = coerce_str(payload.get("city") or query.city)
+    if not city:
+        return False
+    normalized_city = normalize_text(city)
+    normalized_name = normalize_text(query.name)
+    if not normalized_city or not normalized_name or normalized_city not in normalized_name:
+        return False
+    return not any(_evidence_has_location_assertion(item, city) for item in evidence_items)
+
+
 def _normalize_llm_precision(value: Any) -> str:
     text = coerce_str(value).lower()
     return LLM_PRECISION_ALIASES.get(text, text)
@@ -588,6 +679,7 @@ def _weak_evidence_penalty(
     payload: dict[str, Any],
     evidence_items: list[dict[str, Any]],
     *,
+    query: LocationQuery,
     entity_match: float,
     geography_match: float,
     conflict_penalty: float,
@@ -598,6 +690,8 @@ def _weak_evidence_penalty(
         penalty += 0.20
     if any(_truthy_evidence_field(item, "ambiguous") for item in evidence_items):
         penalty += 0.15
+    if _city_is_unsupported_name_hint(payload, query, evidence_items):
+        penalty += 0.10
     if conflict_penalty == 0.0 and entity_match > 0 and geography_match >= 0.20:
         return min(penalty, 0.15)
     return min(penalty, 0.30)
@@ -620,6 +714,7 @@ def _score_llm_location_payload(
     weak_evidence_penalty = _weak_evidence_penalty(
         payload,
         evidence_items,
+        query=query,
         entity_match=entity_match,
         geography_match=geography_match,
         conflict_penalty=conflict_penalty,
@@ -636,6 +731,8 @@ def _score_llm_location_payload(
         - weak_evidence_penalty
     )
     score = min(max(score, 0.0), 1.0)
+    if _city_is_unsupported_name_hint(payload, query, evidence_items):
+        score = min(score, 0.54)
     summary = (
         f"combined={score:.2f}; model={model_confidence:.2f}; "
         f"source={source_quality:.2f}; entity={entity_match:.2f}; "
@@ -847,15 +944,43 @@ async def collect_location_search_evidence(
 ) -> LocationSearchEvidenceResult:
     search_query = _location_search_query(query, entity_type)
     attempt = f"web_search:{entity_type}:{search_query}"
+    logger.info_event(
+        "Collecting location search evidence",
+        event="location.factcheck.web_search.start",
+        context={
+            "entity_type": entity_type,
+            "search_query": search_query,
+            "location_query": _query_context(query),
+            "max_results": max_results,
+        },
+    )
     try:
         evidence = await web_search_client.search(search_query, max_results=max_results)
     except WebSearchError as exc:
+        logger.warning_event(
+            "Location search evidence failed",
+            event="location.factcheck.web_search.failed",
+            context={
+                "entity_type": entity_type,
+                "search_query": search_query,
+                "error": str(exc),
+            },
+        )
         return LocationSearchEvidenceResult(
             evidence=[],
             attempted_queries=[attempt],
             failure_reason=f"WebSearch location evidence failed: {exc}",
         )
     except Exception as exc:
+        logger.warning_event(
+            "Location search evidence unavailable",
+            event="location.factcheck.web_search.unavailable",
+            context={
+                "entity_type": entity_type,
+                "search_query": search_query,
+                "error": str(exc),
+            },
+        )
         return LocationSearchEvidenceResult(
             evidence=[],
             attempted_queries=[attempt],
@@ -863,11 +988,29 @@ async def collect_location_search_evidence(
         )
     normalized = normalize_search_evidence(evidence, limit=max_results)
     if not normalized:
+        logger.warning_event(
+            "Location search returned no usable evidence",
+            event="location.factcheck.web_search.empty",
+            context={
+                "entity_type": entity_type,
+                "search_query": search_query,
+            },
+        )
         return LocationSearchEvidenceResult(
             evidence=[],
             attempted_queries=[attempt],
             failure_reason="WebSearch returned no usable location evidence.",
         )
+    logger.info_event(
+        "Collected location search evidence",
+        event="location.factcheck.web_search.result",
+        context={
+            "entity_type": entity_type,
+            "search_query": search_query,
+            "evidence_count": len(normalized),
+            "evidence": _summarize_search_evidence(normalized),
+        },
+    )
     return LocationSearchEvidenceResult(
         evidence=normalized,
         attempted_queries=[attempt],
@@ -947,6 +1090,15 @@ async def collect_llm_location_fallback_candidate(
     """
     attempt = f"llm_factcheck:{entity_type}:{coerce_str(query.name) or 'unknown'}"
     if search_evidence is not None and not search_evidence:
+        logger.warning_event(
+            "Skipping LLM location factcheck because search evidence is empty",
+            event="location.factcheck.llm.skipped_no_evidence",
+            context={
+                "entity_type": entity_type,
+                "attempt": attempt,
+                "location_query": _query_context(query),
+            },
+        )
         return LocationLLMFallbackResult(
             candidates=[],
             attempted_queries=[attempt],
@@ -986,20 +1138,66 @@ async def collect_llm_location_fallback_candidate(
             "Return evidence as objects when possible, including source, url, source_type, and entity_match.",
             "Include source names or URLs in evidence when known. The backend will recompute the final confidence from model confidence plus evidence quality.",
             "If search_evidence is provided, use only that evidence as factual support.",
+            "Do not treat a website footer, office address, publisher address, or contact address as the entity's physical location.",
+            "If the entity name contains a city name, do not choose that city unless evidence explicitly says the entity/facility/supercomputer is located, hosted, built, deployed, or installed there.",
             "Prefer the facility/site if known; otherwise use the best supported city.",
         ],
     )
+    logger.info_event(
+        "Sending location factcheck request to LLM",
+        event="location.factcheck.llm.request",
+        context={
+            "entity_type": entity_type,
+            "attempt": attempt,
+            "title": request.title,
+            "objective": request.objective,
+            "location_query": request.context.get("location_query"),
+            "observations": request.observations,
+            "constraints": request.constraints,
+            "search_evidence_count": len(search_evidence or []),
+            "search_evidence": _summarize_search_evidence(search_evidence),
+        },
+    )
     try:
         response = await provider_client.analyze(request)
     except Exception as exc:
+        logger.warning_event(
+            "LLM location factcheck failed",
+            event="location.factcheck.llm.failed",
+            context={
+                "entity_type": entity_type,
+                "attempt": attempt,
+                "error": str(exc),
+            },
+        )
         return LocationLLMFallbackResult(
             candidates=[],
             attempted_queries=[attempt],
             failure_reason=f"LLM location factcheck failed: {exc}",
         )
 
+    logger.info_event(
+        "Received location factcheck response from LLM",
+        event="location.factcheck.llm.response",
+        context={
+            "entity_type": entity_type,
+            "attempt": attempt,
+            "provider": response.provider,
+            "model": response.model,
+            "content": _truncate_log_text(response.content, 2000),
+        },
+    )
     payload = _first_json_object(response.content)
     if payload is None:
+        logger.warning_event(
+            "LLM location factcheck response was not strict JSON; attempting repair",
+            event="location.factcheck.llm.non_json",
+            context={
+                "entity_type": entity_type,
+                "attempt": attempt,
+                "content": _truncate_log_text(response.content, 1200),
+            },
+        )
         payload = await _repair_location_payload_from_text(
             provider_client=provider_client,
             raw_text=response.content,
@@ -1009,9 +1207,17 @@ async def collect_llm_location_fallback_candidate(
         )
     if payload is None:
         payload = _payload_from_free_text(response.content, query=query)
-    if payload is None:
+    if payload is None and entity_type != "compute_center":
         payload = _payload_from_query_name_geocode(query)
     if payload is None:
+        logger.warning_event(
+            "LLM location factcheck produced no parseable payload",
+            event="location.factcheck.llm.unparseable",
+            context={
+                "entity_type": entity_type,
+                "attempt": attempt,
+            },
+        )
         return LocationLLMFallbackResult(
             candidates=[],
             attempted_queries=[attempt],
@@ -1032,7 +1238,7 @@ async def collect_llm_location_fallback_candidate(
                     "url": item.get("url"),
                     "text": item.get("snippet") or item.get("content"),
                     "source_type": "web_search",
-                    "entity_match": True,
+                    "entity_match": _search_evidence_entity_match(item, query),
                 }
                 for item in search_evidence
                 if isinstance(item, dict)
@@ -1054,6 +1260,18 @@ async def collect_llm_location_fallback_candidate(
     if candidate is None:
         if city_geocode_failure and rejection_reason == "missing, invalid, or zero latitude/longitude":
             rejection_reason = f"{rejection_reason}; {city_geocode_failure}"
+        logger.warning_event(
+            "Rejected LLM location factcheck candidate",
+            event="location.factcheck.llm.rejected",
+            context={
+                "entity_type": entity_type,
+                "attempt": attempt,
+                "reason": rejection_reason,
+                "payload": payload,
+                "search_evidence_count": len(search_evidence or []),
+                "search_evidence": _summarize_search_evidence(search_evidence),
+            },
+        )
         return LocationLLMFallbackResult(
             candidates=[],
             attempted_queries=[attempt],
@@ -1062,6 +1280,18 @@ async def collect_llm_location_fallback_candidate(
                 + (f": {rejection_reason}." if rejection_reason else ".")
             ),
         )
+    logger.info_event(
+        "Accepted LLM location factcheck candidate",
+        event="location.factcheck.llm.accepted",
+        context={
+            "entity_type": entity_type,
+            "attempt": attempt,
+            "candidate": candidate.to_dict(),
+            "payload": payload,
+            "search_evidence_count": len(search_evidence or []),
+            "search_evidence": _summarize_search_evidence(search_evidence),
+        },
+    )
     return LocationLLMFallbackResult(
         candidates=[candidate],
         attempted_queries=[attempt],