gbanyan/novelty-seeking
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: Improve expert diversity and description reliability

Browse Source 
- Add random seed and diversity hints to expert generation prompt
- Explicitly avoid common professions (醫生、工程師、教師、律師等)
- Change description generation from batch to one-by-one for reliability
- Increase default temperature from 0.7 to 0.95 for more creative output
- Add description_progress SSE event for real-time feedback

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
gbanyan
2025-12-04 11:24:03 +08:00
parent 9079f7a8a9
commit baea210109
3 changed files with 68 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
backend/app/routers/expert_transformation.py
View File

@@ -17,7 +17,7 @@ from ..models.schemas import (
from ..prompts.expert_transformation_prompt import (
get_expert_generation_prompt,
get_expert_keyword_generation_prompt,
get_expert_batch_description_prompt,
get_single_description_prompt,
)
from ..services.llm_service import ollama_provider, extract_json_from_response
@@ -119,34 +119,48 @@ async def generate_expert_transformation_events(
yield f"event: error\ndata: {json.dumps({'error': '無法生成關鍵字'}, ensure_ascii=False)}\n\n"
return
# ========== Step 2: Generate descriptions for each expert keyword ==========
yield f"event: description_start\ndata: {json.dumps({'message': '為專家關鍵字生成創新應用描述...'}, ensure_ascii=False)}\n\n"
# ========== Step 2: Generate descriptions one by one ==========
yield f"event: description_start\ndata: {json.dumps({'message': '為專家關鍵字生成創新應用描述...', 'total': len(all_expert_keywords)}, ensure_ascii=False)}\n\n"
descriptions: List[ExpertTransformationDescription] = []
try:
desc_prompt = get_expert_batch_description_prompt(
query=request.query,
category=request.category,
expert_keywords=[kw.model_dump() for kw in all_expert_keywords]
)
logger.info(f"Description prompt: {desc_prompt[:300]}")
# Build expert lookup for domain info
expert_lookup = {exp.id: exp for exp in experts}
desc_response = await ollama_provider.generate(
desc_prompt, model=model, temperature=temperature
)
logger.info(f"Description response: {desc_response[:500]}")
for idx, kw in enumerate(all_expert_keywords):
try:
expert = expert_lookup.get(kw.expert_id)
expert_domain = expert.domain if expert else ""
desc_data = extract_json_from_response(desc_response)
descriptions_raw = desc_data.get("descriptions", [])
desc_prompt = get_single_description_prompt(
query=request.query,
keyword=kw.keyword,
expert_id=kw.expert_id,
expert_name=kw.expert_name,
expert_domain=expert_domain
)
for desc in descriptions_raw:
if isinstance(desc, dict) and all(k in desc for k in ["keyword", "expert_id", "expert_name", "description"]):
descriptions.append(ExpertTransformationDescription(**desc))
desc_response = await ollama_provider.generate(
desc_prompt, model=model, temperature=temperature
)
except Exception as e:
logger.warning(f"Failed to generate descriptions: {e}")
# Continue without descriptions - at least we have keywords
desc_data = extract_json_from_response(desc_response)
desc_text = desc_data.get("description", "")
if desc_text:
descriptions.append(ExpertTransformationDescription(
keyword=kw.keyword,
expert_id=kw.expert_id,
expert_name=kw.expert_name,
description=desc_text
))
# Send progress update
yield f"event: description_progress\ndata: {json.dumps({'current': idx + 1, 'total': len(all_expert_keywords), 'keyword': kw.keyword}, ensure_ascii=False)}\n\n"
except Exception as e:
logger.warning(f"Failed to generate description for '{kw.keyword}': {e}")
# Continue with next keyword
yield f"event: description_complete\ndata: {json.dumps({'count': len(descriptions)}, ensure_ascii=False)}\n\n"