fix(search): stop hallucinated NL attributes from zeroing category results

A small local model returned a full attribute set for "Schulen in meiner Nähe"
(schools), which applyFacets applied globally — filtering every school away and
showing a wall of "could not filter by" tags as a near-invisible map overlay.

- applyFacets now respects each facet's categoryIds (schools support no facets,
  so bogus attrs are ignored) and skips the model-default "no" on multi facets.
- The unmapped-attribute notice drops recognized OSM tag keys (keeps genuine
  free-text qualities), dedupes, and renders as a readable pill, not faint text.
- Prompt: emit only attributes the user explicitly states; empty otherwise.

Author: Medformatik

apps/web/src/components/search/CategoryFilterBar.tsx

@@ -33,6 +33,27 @@ import { CategoryFiltersPanel } from "./CategoryFiltersPanel";
 import { floatingChipSx, floatingToolbarSx } from "./floatingChipSx";
 import { NlpUnmappedNotice } from "./NlpFilterChips";
 
+// Recognized OSM attribute keys (base form). A small model sometimes echoes
+// these into `unmapped_attributes` instead of leaving them out — they aren't
+// genuinely "unmapped", so they're stripped from the notice.
+const KNOWN_OSM_ATTRIBUTE_KEYS = new Set([
+  "outdoor_seating",
+  "wheelchair",
+  "internet_access",
+  "cuisine",
+  "diet",
+  "takeaway",
+  "delivery",
+  "drive_through",
+  "smoking",
+  "dog",
+  "fee",
+  "payment",
+  "live_music",
+  "organic",
+  "distance",
+]);
+
 // Display order: Mon–Sun; JS day indices
 const DAYS: { key: string; idx: number }[] = [
   { key: "monday", idx: 1 },
@@ -123,11 +144,22 @@ export function CategoryFilterBar() {
   // filter (e.g. "best", "instagrammable") so the user knows they aren't
   // narrowing the results.
   const isNlpActive = useNlpSearchStore((s) => s.isNlpActive);
-  const nlpUnmapped = useNlpSearchStore((s) => s.intent?.unmapped_attributes);
+  const nlpRawUnmapped = useNlpSearchStore((s) => s.intent?.unmapped_attributes);
+  const nlpAttributes = useNlpSearchStore((s) => s.intent?.attributes);
+  // Keep only genuine free-text qualities ("cozy", "best") — drop anything that
+  // is a recognized OSM tag key or already present in `attributes` (small models
+  // duplicate/echo the tag vocabulary here), and dedupe.
+  const nlpUnmapped = useMemo(() => {
+    if (!nlpRawUnmapped) return [];
+    const attrKeys = new Set(Object.keys(nlpAttributes ?? {}));
+    return [
+      ...new Set(
+        nlpRawUnmapped.filter((a) => !attrKeys.has(a) && !KNOWN_OSM_ATTRIBUTE_KEYS.has(a)),
+      ),
+    ];
+  }, [nlpRawUnmapped, nlpAttributes]);
   const unmappedNotice =
-    isNlpActive && nlpUnmapped && nlpUnmapped.length > 0 ? (
-      <NlpUnmappedNotice attributes={nlpUnmapped} />
-    ) : null;
+    isNlpActive && nlpUnmapped.length > 0 ? <NlpUnmappedNotice attributes={nlpUnmapped} /> : null;
   const { rawResults, dominantCategory } = useExploreResults();
   // In text mode there is no active category chip — reuse the facets of the
   // category that the results predominantly belong to (e.g. "McDonald's" →

apps/web/src/components/search/NlpFilterChips.tsx

@@ -17,11 +17,19 @@ export function NlpUnmappedNotice({ attributes }: { attributes: string[] }) {
   return (
     <Box
       sx={{
-        display: "flex",
+        display: "inline-flex",
         alignItems: "center",
         gap: 0.75,
         color: "text.secondary",
         pointerEvents: "auto",
+        // Give the notice an opaque pill background — it floats over the map, so
+        // without this the muted caption is nearly unreadable against the tiles.
+        bgcolor: "background.paper",
+        px: 1.25,
+        py: 0.5,
+        borderRadius: 2,
+        boxShadow: 2,
+        maxWidth: "100%",
       }}
     >
       <InfoOutlinedIcon sx={{ fontSize: 16, flexShrink: 0 }} />

integrations/search-nlp/prompt.ts

@@ -9,11 +9,11 @@ OSM ATTRIBUTE TAGS (use these exact keys in "attributes", string values only):
 outdoor_seating ("yes"/"no"), wheelchair ("yes"/"limited"/"no"), internet_access ("wlan"/"yes"/"no"), cuisine (lowercase, e.g. "italian"), diet:vegan ("yes"/"only"), diet:vegetarian ("yes"/"only"), diet:halal ("yes"), diet:kosher ("yes"), diet:gluten_free ("yes"), takeaway ("yes"/"no"), delivery ("yes"/"no"), drive_through ("yes"/"no"), smoking ("yes"/"no"/"outside"), dog ("yes"/"no"), fee ("yes"/"no"), payment:credit_cards ("yes"/"no"), live_music ("yes"), organic ("yes")
 
 RULES:
-1. Map intent to one or more categories from the list.
-2. Extract OSM attribute filters using only the keys above; values are strings.
+1. Map intent to one or more categories from the list, using the exact IDs above (e.g. "kindergartens", not "kindergarten").
+2. Extract OSM attribute filters using only the keys above; values are strings. Include ONLY attributes the user EXPLICITLY states. If the user mentions no attributes, "attributes" MUST be an empty object {}. Never invent attributes, never fill in defaults, and never emit the value "no" unless the user explicitly excludes that feature (e.g. "no outdoor seating").
 3. "near <place>" => spatial_constraint {type:"near_place", place_name}. "near me" => {type:"current_view"} (the client resolves the user's location).
 4. Time phrases => time_constraint (open_now / open_24h / open_at with day "Monday".."Sunday" and "HH:MM").
-5. Qualities with no OSM tag (quiet, cozy, cheap, family-friendly) go in unmapped_attributes.
+5. Free-text qualities the user typed that have no OSM tag (quiet, cozy, cheap, family-friendly) go in unmapped_attributes. Put ONLY such words here — never OSM tag keys — and leave it empty if there are none.
 6. confidence 0-1: 0.9+ clear intent, 0.5-0.8 ambiguous, below 0.4 if it looks like a place name/address (then categories=[]).
 7. explanation: short human-readable summary.
 8. sort_by defaults to "relevance"; "nearest/closest" => "distance"; "best/top-rated" => "rating".

packages/core/src/stores/__tests__/nlpSearchStore.test.ts

@@ -132,6 +132,30 @@ describe("useNlpSearchStore", () => {
         useNlpSearchStore.getState().activate(intent, bbox, "openai");
         expect(useCategoryFacetStore.getState().selections.wheelchairAccessible).toBeUndefined();
       });
+
+      it("ignores facets not scoped to the active category (e.g. food attrs on a schools search)", () => {
+        // Reproduces the Aachen bug: a small model hallucinated a full attribute
+        // set for "Schulen in meiner Nähe"; none of these apply to schools, so
+        // no facet should be set and the schools results must not be filtered.
+        const intent = makeIntent({
+          categories: ["schools"],
+          attributes: {
+            outdoor_seating: "no",
+            wheelchair: "limited",
+            internet_access: "wlan",
+            cuisine: "no",
+            "diet:vegan": "yes",
+          },
+        });
+        useNlpSearchStore.getState().activate(intent, bbox, "openai");
+        expect(Object.keys(useCategoryFacetStore.getState().selections)).toHaveLength(0);
+      });
+
+      it("skips a multi facet whose value is the model default 'no'", () => {
+        const intent = makeIntent({ categories: ["restaurants"], attributes: { cuisine: "no" } });
+        useNlpSearchStore.getState().activate(intent, bbox, "openai");
+        expect(useCategoryFacetStore.getState().selections.cuisine).toBeUndefined();
+      });
     });
   });
 

packages/core/src/stores/nlpSearchStore.ts

@@ -44,14 +44,23 @@ function applyTimeConstraint(intent: SearchIntent): void {
 function applyFacets(intent: SearchIntent): void {
   const facets = useCategoryFacetStore.getState();
   const attrs = intent.attributes;
+  // Only apply facets that actually belong to the category being searched.
+  // Models (especially small local ones) can emit attributes the user never
+  // asked for; without this scope a hallucinated food/dietary attribute would
+  // get applied to e.g. a "schools" search and filter every result away.
+  const activeCategory = intent.categories[0];
+  if (!activeCategory) return;
   for (const facet of CATEGORY_FACETS) {
+    if (!facet.categoryIds.has(activeCategory)) continue;
     const value = attrs[facet.tag];
     if (value === undefined) continue;
     if (facet.type === "toggle") {
       if ((facet.matchValues ?? []).includes(value)) {
         facets.setMultiFacet(facet.id, ["on"]);
       }
-    } else {
+    } else if (value !== "no") {
+      // Multi facets (e.g. cuisine) take the value verbatim — skip the model's
+      // default "no", which would filter to a nonexistent value (zero results).
       facets.setMultiFacet(facet.id, [value]);
     }
   }