-- Configurable evidence sources — new Settings → Advanced tab lets you pin the number of retrieved evidence snippets per query (4–30). Local/Ollama users who were seeing only 4 sources can now increase this independently of the model's context window. Default remains Auto (scaled dynamically).
-- Multilingual library support — three new embedding models available in Library settings: BAAI/bge-m3 (100+ languages, ~2.3 GB), BAAI/bge-large-zh-v1.5 (Chinese-optimised, ~1.3 GB), and paraphrase-multilingual-MiniLM-L12-v2 (50+ languages, ~470 MB). Download on first use, no extra dependencies. Fixes BM25 tokenization for Chinese, Japanese, and Korean text, which was producing near-zero retrieval scores.
-- Follow-up query regression fixed — on turn 2+, fresh retrieval evidence was silently discarded, causing GPT-4o and similar models to generate a new question rather than an answer. Evidence is now correctly injected into every follow-up turn.
-- Conversation history no longer drops first-turn evidence — the context trimmer now pins the first exchange (which carries the bulk of library evidence) and only evicts middle turns when the budget is exceeded.
-- Query condenser reliability — fixed four bugs: false positives on short comparison questions, missed anaphora detection when queries end with ?, condensation firing before any assistant reply exists, and a hallucination guard that falls back to the original query when the condensed version drifts too far.
-- Security — 12 dependency updates including starlette 0.49.3 → 1.0.1 (malformed Host header CVE), fastapi 0.121.2 → 0.134.0, axios 1.13.2 → 1.17.0, react-router-dom 7.9.6 → 7.17.0, and several Python package patches.