inovel-agent

A conversational agent over a private corpus of 7,655 Chinese web novels (plain TXT, ~36 GB). LLM-built book cards and a Qwen3 multi-vector index turn the raw text into searchable artefacts; the agent searches, verifies constraints per book, recommends with reasons and answers in-book questions. Two-stage evaluation: index supply, agent protocol.

书卡（src/retrieval/cards.py、card_schema.py）。LLM 读整份 digest，按封闭词表填三层字段：题材 12 个一级、41 个二级，模型只填二级、一级由程序推出；元素 40 个，每个必须附一条 ≤20 字的原文摘录，程序在 digest 里核对，核不上的降为「待核」；风格五个标尺（主角结构、爽度、基调、感情线、主角起点）。词表参考起点、纵横的分类体系逐条审定，词表外的词一律丢弃并记录原因。7,655 本出了 7,410 张卡。设计过程见 docs/card-schema-v3.md、docs/book-cards-pilot.md。