v1.4.0 — Local models were only seeing part of the prompt
When the app talks to a local Ollama model, it never told the model how much context to use — so Ollama applied its 4096-token default and silently dropped the beginning of every prompt that exceeded it. No error, no warning: the response simply came back written from a partial view.
Fixed
Cross-document summaries were losing the instructions and the first document. The prompt for a unified or comparison summary is assembled as instructions, then each document's block. Measured: 35% of it was being discarded from the front — which meant the model often never saw what it was being asked to do, and never saw the first document at all, while the instructions it did receive told it to cite every claim. The result could cite a document it had not read. This path is now given the context window it needs.
Q&A lost the opening of the summary it was reasoning from. Only 2% was cut, but truncation removes the front of the prompt first, and that is where the summary excerpt sits. The question itself and recent turns survived at the end, so answers stayed on topic while quietly missing their opening context.
Summaries were running at the very edge of the window. A standard chunk occupied 96% of the 4096-token context once the model's own output was counted, leaving room for only a short answer. Longer summaries were being cut off.
A stale index build could clear the new document's search index. Switching documents while an index was being built could leave the new document marked as failed, which in turn stopped its session from being saved with an index.
A tab you closed could come back, and a document you had just opened could be silently replaced by a slower session restore finishing behind it.
Summaries made from part of a document now say so. A page-range summary, a custom-template summary that hit the length budget, and a collection member summarised from its opening pages all used to be stored exactly like a complete summary — and reopening the document showed them as such. They now carry a note in the text itself.
Copy confirmation could appear off-screen. When a banner was showing, the document view overflowed past the bottom of the window and took the confirmation with it. The layout no longer overflows.
Also
electron 44.1.1. Cross-document summary output no longer leaves page citations that jump to the wrong document. Deleting a session no longer risks leaving an unrecoverable phantom entry in the index.
Internal: 2,456 unit tests / 117 files. A four-agent audit of this release found five high-severity issues — four of them introduced by the fixes above — and all are addressed here.
🇰🇷 한국어
로컬 Ollama 모델을 쓸 때 앱이 컨텍스트 크기를 알려 주지 않아, Ollama 의 기본값 4096 토큰이 적용되고 그것을 넘는 프롬프트는 앞부분부터 조용히 버려지고 있었습니다. 에러도 경고도 없이, 일부만 보고 쓴 답이 그대로 돌아왔습니다.
수정
교차 문서 요약이 지시문과 첫 문서를 잃고 있었습니다. 통합·비교 요약의 프롬프트는 지시문 다음에 각 문서 블록 순서로 조립됩니다. 실측 결과 그 앞부분 35%가 버려지고 있었습니다 — 모델이 무엇을 하라는 것인지조차 못 본 채, 첫 문서도 못 본 채 요약했다는 뜻입니다. 그러면서 전달된 지시문은 "모든 근거에 출처를 표기하라"고 요구했기 때문에, 읽지도 않은 문서를 인용하는 일이 생길 수 있었습니다. 이제 이 경로가 필요한 만큼의 컨텍스트를 받습니다.
Q&A 가 근거로 삼는 요약의 앞부분을 잃고 있었습니다. 잘린 양은 2%뿐이지만 절단은 프롬프트 앞에서부터 일어나고, 거기가 요약 발췌가 놓이는 자리입니다. 질문과 최근 대화는 뒤에 있어 살아남았기 때문에 답변은 주제에 맞되 도입부 맥락만 조용히 빠져 있었습니다.
요약이 컨텍스트 한계 바로 앞에서 돌고 있었습니다. 기본 청크 하나가 모델의 출력까지 합치면 4096 토큰 창의 96%를 썼습니다 — 짧은 답변만 들어갈 여유였고, 조금만 길어지면 잘렸습니다.
낡은 색인 작업이 새 문서의 검색 인덱스를 지울 수 있었습니다. 색인 중 문서를 전환하면 새 문서가 실패 표시된 채 남고, 그 결과 인덱스 없이 세션이 저장되는 경우가 있었습니다.
닫은 탭이 되살아나거나, 방금 연 문서가 뒤늦게 끝난 세션 복원에 조용히 덮어써질 수 있었습니다.
문서의 일부만으로 만든 요약에 그 사실이 표시됩니다. 페이지 범위 요약, 길이 예산에 걸린 커스텀 템플릿 요약, 앞부분만 읽고 만든 컬렉션 멤버 요약 — 셋 다 완전한 요약과 똑같이 저장돼 재오픈하면 구분할 수 없었습니다. 이제 본문에 안내가 남습니다.
복사 확인이 화면 밖에 뜰 수 있었습니다. 배너가 떠 있으면 문서 뷰가 창 아래로 넘치면서 확인 알림도 함께 밀려났습니다. 레이아웃이 더 이상 넘치지 않습니다.
그 외
electron 44.1.1. 교차 문서 요약 결과에 남던 페이지 인용이 엉뚱한 문서로 점프하지 않습니다. 세션 삭제가 회수 불가능한 유령 항목을 남길 위험을 없앴습니다.
내부: 단위 테스트 2,456건 / 117파일. 이번 릴리즈를 대상으로 한 4-에이전트 감사에서 높은 심각도 5건이 나왔고 그중 4건이 위 수정들이 만든 것이었습니다 — 전부 이번에 함께 처리했습니다.
Windows installer SHA-256
fef60c0cd8e339945ef6b3f3d30677fffe4266dd3cabcfc4150dd4172f6b7577 Local-PDF-Analyzer-Setup-1.4.0.exe