feat(index): Add fallback for community report extraction on provider…

[Luotianyi-0712-tech]

Problem

When I first used and configured the deepseek api, such an error occurred“Pipeline error: 'community'”,The reason is that GraphRAG's community report generation forces the use of JSON mode, which is incompatible with the DeepSeek API.
The create_community_reports workflow fails entirely when using LLM providers that do not support response_format with Pydantic models (i.e., json_schema mode). This includes popular providers like DeepSeek, which only supports {"type": "json_object"} and returns:BadRequestError: This response_format type is unavailable now
When all community report requests fail, the pipeline crashes with KeyError: 'community' because community_reports DataFrame is empty.

Solution

Implement a three-tier fallback strategy in CommunityReportsExtractor:

1. Primary: Try structured output with Pydantic response_format (best quality, works on OpenAI/Azure/Anthropic).
2. Fallback 1: If provider rejects json_schema, catch the error and retry with {"type": "json_object"} (works on DeepSeek, Gemini 1.5, etc.).
3. Fallback 2: If even json_object is rejected, fall back to plain text completion and extract JSON manually via regex (works on Ollama and other minimal providers).

Provider Support Matrix

Provider	json_schema (Pydantic)	json_object	Plain Text
OpenAI / Azure	✅	✅	✅
Anthropic	✅	✅	✅
Gemini 2.0+	✅	✅	✅
Gemini 1.5	❌ (litellm client validation)	✅	✅
DeepSeek	❌	✅	✅
Ollama	partial	partial	✅

Changes

• community_reports_extractor.py: Added _is_unsupported_response_format_error(), _parse_json_from_text(), and three-tier try/catch logic.
• No breaking changes. Existing users on OpenAI/Azure continue to use json_schema without any behavior change.

Testing

• Tested with OpenAI (gpt-4o-mini) — uses json_schema path
• Tested with DeepSeek (deepseek-v4-flash) — falls back to json_object path

[Luotianyi-0712-tech]

@microsoft-github-policy-service agree