Problem

KnowledgeExtractorAgent passes response_format={"type": "json_object"} to the LLM. DeepSeek API does not support this parameter and returns JSON wrapped in markdown code fences (```json ... ```), causing json.loads() to fail.

Steps to Reproduce

1. Configure LLM to use DeepSeek (LLM_BINDING=deepseek, LLM_MODEL=deepseek-chat)
2. Upload a document and trigger knowledge point extraction
3. Extraction fails with JSON decode error

Expected

The extractor should handle markdown-wrapped JSON responses gracefully.

Suggested Fix

In deeptutor/agents/knowledge_extractor/extractor_agent.py, strip markdown fences before parsing:

cleaned = response.strip()
if cleaned.startswith("```"):
    first_newline = cleaned.find("\n")
    if first_newline != -1:
        cleaned = cleaned[first_newline + 1:]
    if cleaned.rstrip().endswith("```"):
        cleaned = cleaned.rstrip()[:-3].rstrip()
result = json.loads(cleaned)

Also consider removing response_format={"type": "json_object"} for providers that don't support it, or making it configurable.

Environment

• LLM: DeepSeek (deepseek-chat)
• Python 3.14, Windows 11