Merge pull request #19 from brianmeyer/codex/rec-192-193-tag-followups

Harden generated media tags and batch tag merge

Author: brianmeyer

src/recallforge/search.py

@@ -1530,6 +1530,18 @@ def search_batch(
     if not batch_queries:
         return []
 
+    def _merge_tags(items: List[Any]) -> Optional[List[str]]:
+        merged: List[str] = []
+        seen: set[str] = set()
+        for item in items:
+            for tag in getattr(item, "tags", None) or []:
+                cleaned = str(tag or "").strip().lower()
+                if not cleaned or cleaned in seen:
+                    continue
+                seen.add(cleaned)
+                merged.append(cleaned)
+        return merged or None
+
     def run_single_query(q: BatchQuery) -> List[tuple]:
         """Run a single query and return (result, score) tuples."""
         mode = q.mode or "hybrid"
@@ -1582,11 +1594,14 @@ def run_single_query(q: BatchQuery) -> List[tuple]:
             if filepath not in merged:
                 merged[filepath] = {
                     'result': result,
+                    'results': [result],
                     'rrf_score': 0.0,
                     'query_indices': set(),
                     'query_scores': {},
                     'best_score': 0.0,
                 }
+            else:
+                merged[filepath]['results'].append(result)
 
             # RRF contribution: rank-based, not insertion-order-based
             merged[filepath]['rrf_score'] += weight / (rrf_k + rank + 1)
@@ -1612,7 +1627,7 @@ def run_single_query(q: BatchQuery) -> List[tuple]:
             score=data['rrf_score'],
             source=','.join(str(i) for i in sorted(data['query_indices'])),
             query_scores=data['query_scores'],
-            tags=getattr(result, "tags", None),
+            tags=_merge_tags(data['results']),
         ))
 
     final_results.sort(key=lambda x: x.score, reverse=True)

src/recallforge/storage/indexing_ops.py

@@ -132,6 +132,10 @@ def _parse_generated_media_tags(self, raw: str) -> List[str]:
         if not text:
             return []
 
+        fenced_match = re.match(r"^```(?:[A-Za-z0-9_+-]+)?\s*\n?(.*?)\n?```$", text, flags=re.DOTALL)
+        if fenced_match:
+            text = fenced_match.group(1).strip()
+
         candidates: List[str] = []
         if text.startswith("[") and text.endswith("]"):
             try:

tests/test_search_batch.py

@@ -525,6 +525,67 @@ def mock_search(self, query):
                 self.assertIn(0, results[0].query_scores)
                 self.assertIn(1, results[0].query_scores)
 
+    def test_same_document_merges_tags_deterministically(self):
+        """Duplicate hits should merge tag sets in stable first-seen order."""
+        backend = StubBackend()
+        storage = StubStorage()
+
+        results_list = [
+            [
+                type('HybridResult', (), {
+                    'filepath': 'shared.md',
+                    'display_path': 'shared.md',
+                    'title': 'shared.md',
+                    'context': None,
+                    'hash': 'h1',
+                    'docid': 'd1',
+                    'collection': 'test',
+                    'modified_at': '2026-01-01',
+                    'body_length': 100,
+                    'body': 'shared content',
+                    'score': 0.8,
+                    'source': 'hybrid',
+                    'tags': ['alpha', 'shared'],
+                }),
+            ],
+            [
+                type('HybridResult', (), {
+                    'filepath': 'shared.md',
+                    'display_path': 'shared.md',
+                    'title': 'shared.md',
+                    'context': None,
+                    'hash': 'h1',
+                    'docid': 'd1',
+                    'collection': 'test',
+                    'modified_at': '2026-01-01',
+                    'body_length': 100,
+                    'body': 'shared content',
+                    'score': 0.9,
+                    'source': 'hybrid',
+                    'tags': ['shared', 'beta'],
+                }),
+            ],
+        ]
+
+        call_idx = [0]
+
+        def mock_search(self, query):
+            idx = call_idx[0]
+            call_idx[0] += 1
+            return results_list[idx]
+
+        with patch.object(HybridSearcher, '__init__', lambda self, **kwargs: None):
+            with patch.object(HybridSearcher, 'search', mock_search):
+                results = search_batch(
+                    ["query one", "query two"],
+                    backend=backend,
+                    storage=storage,
+                    limit=10,
+                )
+
+                self.assertEqual(len(results), 1)
+                self.assertEqual(results[0].tags, ["alpha", "shared", "beta"])
+
 
 if __name__ == "__main__":
-    unittest.main()
+    unittest.main()

tests/test_storage.py

@@ -1340,6 +1340,28 @@ def test_memory_lookup_surfaces_media_tags(self):
             ["neural network", "diagram", "hidden layers"],
         )
 
+    def test_generated_media_tags_strip_fenced_json(self):
+        embedder = CaptioningEmbedder()
+
+        def fenced_json(_prompt: str, max_tokens: int = 60) -> str:
+            return '```json\n["diagram", "hidden layers", "neural network"]\n```'
+
+        embedder.generate_text = fenced_json
+
+        self.backend.index_image(
+            path=self.image_path,
+            collection="test",
+            embed_func=embedder,
+            caption_media=True,
+        )
+
+        rows = self.backend._embeddings_table.search().where("content_type = 'image'").to_list()
+        self.assertEqual(len(rows), 1)
+        self.assertEqual(
+            json.loads(rows[0].get("tags") or "[]"),
+            ["diagram", "hidden layers", "neural network"],
+        )
+
     def test_index_video_keeps_parent_memory_and_links_children(self):
         embedder = CaptioningEmbedder()
         logical_path = str(Path(self.video_path).expanduser().resolve())