Clean extracted schema json

src/neo4j_graphrag/experimental/components/schema.py

@@ -15,6 +15,7 @@
 from __future__ import annotations
 
 import json
+import re
 
 import neo4j
 import logging
@@ -554,6 +555,15 @@ def _filter_relationships_without_labels(
             relationship_types, "relationship type"
         )
 
+    def _clean_json_content(self, content: str) -> str:
+        content = content.strip()
+
+        # Remove markdown code block markers if present
+        content = re.sub(r'^```(?:json)?\s*', '', content, flags=re.MULTILINE)
+        content = re.sub(r'```\s*$', '', content, flags=re.MULTILINE)
+
+        return content.strip()
+
     @validate_call
     async def run(self, text: str, examples: str = "", **kwargs: Any) -> GraphSchema:
         """
@@ -575,6 +585,9 @@ async def run(self, text: str, examples: str = "", **kwargs: Any) -> GraphSchema
             # Re-raise the LLMGenerationError
             raise LLMGenerationError("Failed to generate schema from text") from e
 
+        # Clean response
+        content = self._clean_json_content(content)
+
         try:
             extracted_schema: Dict[str, Any] = json.loads(content)
 

tests/unit/experimental/components/test_schema.py

@@ -960,6 +960,86 @@ async def test_schema_from_text_filters_relationships_without_labels(
     assert ("Person", "MANAGES", "Organization") in schema.patterns
 
 
+@pytest.fixture
+def valid_schema_json_with_markdown() -> str:
+    return """```json
+{
+    "node_types": [
+        {
+            "label": "Person",
+            "properties": [
+                {"name": "name", "type": "STRING"}
+            ]
+        },
+        {
+            "label": "Organization",
+            "properties": [
+                {"name": "name", "type": "STRING"}
+            ]
+        }
+    ],
+    "relationship_types": [
+        {
+            "label": "WORKS_FOR",
+            "properties": [
+                {"name": "since", "type": "DATE"}
+            ]
+        }
+    ],
+    "patterns": [
+        ["Person", "WORKS_FOR", "Organization"]
+    ]
+}
+```"""
+
+
+@pytest.fixture
+def valid_schema_json_with_markdown_no_language() -> str:
+    return """```
+{
+    "node_types": [
+        {
+            "label": "Person",
+            "properties": [
+                {"name": "name", "type": "STRING"}
+            ]
+        }
+    ]
+}
+```"""
+
+
+def test_clean_json_content_markdown_with_json_language(
+    schema_from_text: SchemaFromTextExtractor,
+) -> None:
+    content = """```json
+{"node_types": [{"label": "Person"}]}
+```"""
+
+    cleaned = schema_from_text._clean_json_content(content)
+    assert cleaned == '{"node_types": [{"label": "Person"}]}'
+
+
+def test_clean_json_content_markdown_without_language(
+    schema_from_text: SchemaFromTextExtractor,
+) -> None:
+    content = """```
+{"node_types": [{"label": "Person"}]}
+```"""
+
+    cleaned = schema_from_text._clean_json_content(content)
+    assert cleaned == '{"node_types": [{"label": "Person"}]}'
+
+
+def test_clean_json_content_plain_json(
+    schema_from_text: SchemaFromTextExtractor,
+) -> None:
+    content = '{"node_types": [{"label": "Person"}]}'
+
+    cleaned = schema_from_text._clean_json_content(content)
+    assert cleaned == '{"node_types": [{"label": "Person"}]}'
+
+
 @pytest.mark.asyncio
 @patch("neo4j_graphrag.experimental.components.schema.get_structured_schema")
 async def test_schema_from_existing_graph(mock_get_structured_schema: Mock) -> None: