[BUG: Breaking] Ollama invalid JSON schema in format

[xiaoyao9184]

🧨 Describe the Bug

When using Marker with an Ollama backend and providing a Pydantic v2 schema via SectionHeaderSchema and PageSchema .model_json_schema(), Marker constructs a format schema that discards $defs, causing $ref references inside properties to break. Ollama then returns:

{"error":"invalid JSON schema in format"}

This is a breaking bug, as $ref is required for any schema with nested models (e.g., list of subobjects), and Marker’s current format_schema logic does not preserve the $defs block from the original model_json_schema().

📄 Input Document

N/A — bug is schema-level.

📤 Output Trace / Stack Trace

Click to expand

{"error":"invalid JSON schema in format"}

⚙️ Environment

Please fill in all relevant details:

• Marker version: 1.8.0
• Surya version:
• Python version: 3.11
• PyTorch version:
• Transformers version:
• Operating System (incl. container info if relevant):

✅ Expected Behavior

Add "$defs": schema["$defs"] if "$defs" in schema else {} in https://github.com/datalab-to/marker/blob/edbcb8cf4ae31da9bdf57f0486b558f56e0a1484/marker/services/ollama.py#L39-L44

[kipavy]

https://github.com/datalab-to/marker/issues/907#issuecomment-3384031679