Spaces:
Sleeping
Sleeping
Commit History
-
Fix empty-content crash with reasoning models (GLM-5.2 default)
Reasoning models spend tokens thinking before writing the visible answer, and those tokens count against max_tokens. The seed-keyword step's max_tokens=400 left GLM-5.2 (the default model) no room to answer: the API returned content=None with finish_reason=length, and re.search(None) crashed every default-settings plan generation with 'Unexpected error: expected string or bytes-like object'. - app.py: raise the seed-keyword budget 400 -> 2000 (verified sufficient against the live API with the real prompt). - llm.py chat(): retry once with +3000 headroom when content is empty at finish_reason=length, and raise a clear LLMError instead of returning None if content is still empty - covers every call site centrally. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>d4ab656
-
Load RAG index eagerly at container startup instead of on first request
Verified signaturea9e0a97 -
Initial deploy: full app with 4-tier keyword research, per-task models; RAG index served from a separate private dataset repo
Verified signaturef23046e