Meeting Memory: Ask Your Meetings Questions and Get the Exact Lines Back, Built for My Club Friends
Описание и применение
Meeting Memory позволяет задавать вопросы по записям встреч и получать ответы с точными ссылками на строки, говорящего и временной меткой. Инструмент использует поиск по ключевым словам в SQLite и генерацию ответов с помощью открытых моделей, обеспечивая прозрачность и отсутствие догадок.
Задача
Recurring project meetings bury decisions and action items in long chats, making it hard to recall what was decided.
Результат
On 16 answerable questions, plain keyword search achieved 7 correct answers, while synonym expansion and per‑meeting sampling achieved 7 correct and 7 partly right. Retrieval was the bottleneck; the model was not called when no relevant segments were found. Extraction yielded 44 items (10 decisions, 7 actions, 27 open questions) with varying support levels. Decisions were the strongest result; open questions were weak.
Рабочий процесс
- Load meeting transcripts into the system
- Parse transcripts into segments (speaker, start time, text, id)
- Store segments and keyword index in SQLite
- Extract decisions, actions, and open questions using open-weight models (openai/gpt-oss-120b)
- Index extracted items with quotes
- When a user asks a question, perform keyword search (SQLite full-text, BM25) across all meetings
- If relevant segments found, pass them to answer model (qwen/qwen3.8-27b) to generate answer with citations
- If no relevant segments, return "Not found"
- Display answer with speaker, timestamp, and source lines in the UI
Доказательства
Meeting Memory loads meeting transcripts and lets you ask questions across all of them. Every answer shows the speaker, the timestamp and the lines it came from.
The model answers using only those, with numbered citations. Citations to anything it wasn’t given are removed, an answer with no valid citation is flagged, and if nothing relevant is found the model isn’t called at all.
Answering, 16 answerable questions: Plain keyword search 7 correct, 5 partly right, 4 wrong. Synonyms + per‑meeting sampling 7 correct, 7 partly right, 2 wrong.
Extraction, final meeting, 44 items: Decisions 10 items, 7 fully supported, 3 partly, 0 unsupported. Actions 7 items, 3 fully supported, 1 partly, 3 unsupported. Open questions 27 items, 5 fully supported, 2 partly, 20 unsupported.
- Дата записи: 2026-10-04T11:55:49.496Z