Spaces:
Running on Zero
Running on Zero
Commit History
-
fix(workflows): survive Qwen3.8 thinking-token exhaustion in chat workflows
AndrianBalanescuLive smoke found summarization returning empty summaries post quota reset: the 1zero combo (Qwen3.8 GGUF) narrates a Thinking Process section that burns ~1450 completion tokens, so 320-800 token budgets ended with finish_reason=length and reasoning-only content that strip_reasoning empties. /no_think prefixes/suffixes and chat_template_kwargs.enable_thinking=false do not suppress it (verified direct-to-Space on both Q8_0 and Q4_K_M quants and through the gateway). Fix: - chat budgets raised to 2048 in summarizer and extractor so thinking completes and the answer arrives with finish_reason=stop - anti-CoT system instruction added (nu afisa procesul de gandire), verified to help on some prompts - regression test asserting the summarizer system prompt - workflows/README.md rewritten with real docs (was a malformed @file placeholder) including the verified quirk behavior Live smoke: all 4 workflows PASS (search 0.521, STT 1.0s, extraction 3.6s, summarization 23.3s clean Romanian su…0ffaa55 -
feat(workflows): add OmniRoute summarization, transcription, search, extraction
AndrianBalanescuMilestone 5 line 69: document summarization, meeting transcription, semantic search, and structured extraction workflows over verified OmniRoute routes (1zero chat, whisperfw STT, infinity embeddings). - moldovan-qwen/workflows package with OmniRouteClient (chat/STT/embed) - map-reduce summarizer with empty-content retry and reasoning stripping - meeting transcriber (multi-file, sorted segments) - semantic search index (cosine, dim-1024 validation, save/load) - structured extractor with JSON-block parsing and one correction retry - transport-only single retry in client (no blind fallback replay) - 36 offline unit tests (FakeClient), live smoke script for all 4 - live smoke verified: all 4 PASS against :201287036d8f