self-trained2 / reply_engine.py

Commit History

23 days ago

  1. Neurones Pro 1.5: upgrade neurones_pro.py in place (id neurones-pro-1.5, default everywhere; neurones-pro-1.0 remapped). Fixes the old prompt that hard-capped answers at 100-200 words and banned tables. New: citation-backed automatic web search for time-sensitive questions, two-speed output budget (fast_cap 900 / reply_cap 2400), rewritten system prompt (complete answers, markdown when useful, honest uncertainty, ask-once-when-ambiguous, brief refusals).

    Verified signature
    68b2187
  2. Fix truncated replies + retire Flash server-side. New reply_engine: caps output per request (OTPM protection), fails over to a different model on 429 instead of retrying the exhausted one, and on finish_reason=length feeds the partial answer back and continues it instead of showing a fragment. main.py: uses the engine for both normal and deep-search replies, returns model_used + reply_incomplete, honest 'at capacity' message instead of a vague error, guest memory isolated (no more shared/anonymous history or learned facts), history no longer sliced to 200/250 chars. Added GET /api/models so clients read the model list from the server. models/registry.py honours enabled=False; neurones-flash-2.0 retired (requests fall back to Neurones Pro 1.0).

    Verified signature
    4b8904a