fix(research): SearXNG searches keep the owner's recency window

toolgate/executors/research_web.py

@@ -81,6 +81,25 @@ def _community_snapshot(item: dict, source: str, timeout: float) -> dict:
     return item
 
 
+def _searx_time_range(recency_days: int) -> dict:
+    """SearXNG's narrowest window that still covers recency_days (it knows day/week/month/year)."""
+    for days, name in ((1, "day"), (7, "week"), (31, "month"), (366, "year")):
+        if recency_days <= days:
+            return {"time_range": name}
+    return {}
+
+
+def _older_than(item: dict, recency_days: int) -> bool:
+    """True when the row carries a publication date before the requested window."""
+    try:
+        published = datetime.fromisoformat(str(item.get("published_at")).replace("Z", "+00:00"))
+    except ValueError:
+        return False
+    if published.tzinfo is None:
+        published = published.replace(tzinfo=UTC)
+    return published < datetime.now(UTC) - timedelta(days=recency_days)
+
+
 def _searx(query: str, source: str, limit: int, recency_days: int, timeout: float) -> list[dict]:
     deadline = time.monotonic() + max(1.0, timeout)
     source_domain = SOURCE_DOMAINS[source]
@@ -96,7 +115,7 @@ def _searx(query: str, source: str, limit: int, recency_days: int, timeout: floa
         f"{base_url}/search",
         params={"q": scoped_query, "format": "json", "language": tool_settings.search_language(),
                 "safesearch": tool_settings.safe_search(), "categories": "general",
-                **({"time_range": "day" if recency_days <= 1 else "month"} if recency_days <= 45 else {})},
+                **_searx_time_range(recency_days)},
         headers=headers,
         timeout=timeout,
     )
@@ -107,6 +126,9 @@ def _searx(query: str, source: str, limit: int, recency_days: int, timeout: floa
         if len(results) >= limit:
             break
         item = _normalize(row.get("title"), row.get("url"), row.get("content"), source, row.get("publishedDate"))
+        # SearXNG's windows are coarse (a week for "last 3 days"), so dated rows are checked too.
+        if item and _older_than(item, recency_days):
+            continue
         hostname = urlsplit(item["url"]).hostname.lower() if item and urlsplit(item["url"]).hostname else ""
         if source == "general" and not domain and any(
             hostname == blocked or hostname.endswith(f".{blocked}")

toolgate/tests/test_searx_recency.py

@@ -0,0 +1,34 @@
+"""A SearXNG search keeps the owner's recency window (hunt toolgate-4 F3)."""
+
+from datetime import UTC, datetime, timedelta
+from unittest.mock import Mock, patch
+
+import pytest
+
+from toolgate.core import control_plane
+from toolgate.executors import research_base, research_web
+
+
+@pytest.mark.parametrize(("days", "window"), [
+    (1, "day"), (3, "week"), (7, "week"), (14, "month"), (31, "month"), (90, "year"), (400, None),
+])
+def test_the_narrowest_searxng_window_that_covers_the_request(days, window):
+    assert research_web._searx_time_range(days).get("time_range") == window
+
+
+def test_rows_dated_before_the_window_are_dropped(tmp_path, monkeypatch):
+    monkeypatch.setattr(control_plane, "DB_PATH", tmp_path / "gate.db")
+    now = datetime.now(UTC)
+    answer = Mock(content=b"{}", raise_for_status=Mock())
+    answer.json.return_value = {"results": [
+        {"title": "Old", "url": "https://a.example.com/old", "content": "old",
+         "publishedDate": (now - timedelta(days=20)).isoformat()},
+        {"title": "New", "url": "https://a.example.com/new", "content": "new",
+         "publishedDate": (now - timedelta(days=2)).isoformat()},
+        {"title": "Undated", "url": "https://a.example.com/undated", "content": "undated"},
+    ]}
+    with patch.object(research_base, "_public_https_url", return_value=True), \
+         patch.object(research_web.httpx, "get", return_value=answer) as get:
+        rows = research_web._searx("invoice tools", "general", 5, 7, 5)
+    assert get.call_args.kwargs["params"]["time_range"] == "week"
+    assert [row["title"] for row in rows] == ["New", "Undated"]