From f151c0865e447903c1111fa9ce1d90de36c2c2fe Mon Sep 17 00:00:00 2001
From: vikingowl <christian@nachtigall.dev>
Date: Sat, 25 Apr 2026 17:48:48 +0200
Subject: [PATCH] fix(discovery): use JSON schema instead of JSONMode for LLM
 enricher

Replaces JSONMode:true with an embedded enricher_schema.json so Gemini
returns structured output against a typed schema, preventing empty {} responses.
Adds an all-empty warning when the LLM returns a valid but blank payload.
---
 .../enrich/assets/enricher_schema.json        | 21 ++++++++++++++
 .../domain/discovery/enrich/llm_enricher.go   | 11 ++++++-
 .../discovery/enrich/llm_enricher_test.go     | 29 +++++++++++++++++--
 3 files changed, 58 insertions(+), 3 deletions(-)
 create mode 100644 backend/internal/domain/discovery/enrich/assets/enricher_schema.json

diff --git a/backend/internal/domain/discovery/enrich/assets/enricher_schema.json b/backend/internal/domain/discovery/enrich/assets/enricher_schema.json
new file mode 100644
index 0000000..66c7891
--- /dev/null
+++ b/backend/internal/domain/discovery/enrich/assets/enricher_schema.json
@@ -0,0 +1,21 @@
+{
+  "type": "object",
+  "required": ["category", "opening_hours", "description"],
+  "properties": {
+    "category": {
+      "type": "string",
+      "description": "Market category. Use empty string if unclear.",
+      "enum": ["mittelaltermarkt", "weihnachtsmarkt", "ritterfest",
+               "handwerkermarkt", "schlossfest", "ritterturnier",
+               "kirchweih", ""]
+    },
+    "opening_hours": {
+      "type": "string",
+      "description": "Brief German summary, e.g. Sa-So 10:00-18:00. Empty if unclear."
+    },
+    "description": {
+      "type": "string",
+      "description": "1-3 sentences in German. Empty if insufficient source info."
+    }
+  }
+}
diff --git a/backend/internal/domain/discovery/enrich/llm_enricher.go b/backend/internal/domain/discovery/enrich/llm_enricher.go
index 06abcb0..81e1378 100644
--- a/backend/internal/domain/discovery/enrich/llm_enricher.go
+++ b/backend/internal/domain/discovery/enrich/llm_enricher.go
@@ -2,6 +2,7 @@ package enrich
 
 import (
 	"context"
+	_ "embed"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -12,6 +13,9 @@ import (
 	"marktvogt.de/backend/internal/pkg/ai"
 )
 
+//go:embed assets/enricher_schema.json
+var enricherSchemaJSON []byte
+
 // maxScrapeURLs limits how many quellen we pull per enrichment call.
 // Real-world queue rows rarely have more than 2-3 sources; 5 is a safe cap
 // that prevents a worst-case row from burning minutes on fetches.
@@ -88,7 +92,7 @@ func (e *ProviderLLMEnricher) EnrichMissing(ctx context.Context, req LLMRequest)
 	resp, err := e.AI.Chat(ctx, &ai.ChatRequest{
 		SystemPrompt: systemPrompt,
 		UserMessage:  userPrompt,
-		JSONMode:     true,
+		JSONSchema:   enricherSchemaJSON,
 		Grounded:     true,
 		CallType:     "enrich_b",
 		Temperature:  0.1,
@@ -102,6 +106,11 @@ func (e *ProviderLLMEnricher) EnrichMissing(ctx context.Context, req LLMRequest)
 		return Enrichment{}, fmt.Errorf("parse llm response: %w (content=%q)", err, resp.Content)
 	}
 
+	if parsed.Category == "" && parsed.OpeningHours == "" && parsed.Description == "" {
+		slog.WarnContext(ctx, "llm enrich: all fields empty despite successful call",
+			"markt", req.MarktName, "stadt", req.Stadt)
+	}
+
 	// Build the Enrichment payload with only the fields the model produced.
 	now := time.Now().UTC()
 	out := Enrichment{
diff --git a/backend/internal/domain/discovery/enrich/llm_enricher_test.go b/backend/internal/domain/discovery/enrich/llm_enricher_test.go
index 18b7d71..20a4894 100644
--- a/backend/internal/domain/discovery/enrich/llm_enricher_test.go
+++ b/backend/internal/domain/discovery/enrich/llm_enricher_test.go
@@ -44,8 +44,11 @@ func TestLLMEnricher_HappyPath(t *testing.T) {
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
-	if stub.seen.JSONMode != true {
-		t.Fatalf("JSONMode must be set")
+	if stub.seen.JSONMode {
+		t.Fatal("JSONMode must not be set when JSONSchema is used")
+	}
+	if len(stub.seen.JSONSchema) == 0 {
+		t.Fatal("JSONSchema must be set")
 	}
 
 	// Result carries LLM fields and provenance llm.
@@ -148,6 +151,28 @@ func TestLLMEnricher_EmptyFieldsNoProvenance(t *testing.T) {
 	}
 }
 
+func TestLLMEnricher_SendsJSONSchema(t *testing.T) {
+	stub := &stubProvider{
+		content: `{"category":"mittelaltermarkt","opening_hours":"Sa-So 10-18","description":"Markt."}`,
+	}
+	scraper := &stubScraper{responses: map[string]string{"https://example.com": "some content"}}
+	e := NewLLMEnricher(stub, scraper)
+	_, err := e.EnrichMissing(context.Background(), LLMRequest{
+		MarktName: "Testmarkt",
+		Stadt:     "Dresden",
+		Quellen:   []string{"https://example.com"},
+	})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if stub.seen.JSONMode {
+		t.Error("expected JSONMode to be false when schema is used")
+	}
+	if len(stub.seen.JSONSchema) == 0 {
+		t.Error("expected JSONSchema to be set")
+	}
+}
+
 func TestLLMEnricher_CapsURLsAtFive(t *testing.T) {
 	// Supply 7 URLs; only the first 5 should be fetched.
 	urls := []string{"u1", "u2", "u3", "u4", "u5", "u6", "u7"}