diff --git a/README.md b/README.md index ca52b61..1a4efcc 100644 --- a/README.md +++ b/README.md @@ -43,6 +43,10 @@ config/config.local.json 使用 MYwrite skill，为这个 Zotero 条目生成一条 AI 阅读笔记：SXAIQUJT。 ``` +```text +使用 MYwrite skill，为这个 Zotero 条目生成满血版深度阅读笔记：SXAIQUJT。不要跳过已有笔记，生成一条新的更详细 Zotero 子笔记。 +``` + ```text 使用 MYwrite skill，给这些 Zotero 条目批量生成 AI 阅读笔记：SXAIQUJT、X7GJZ627、ZCZXGRAM。 ``` @@ -65,6 +69,8 @@ config/config.local.json 更稳妥的方式是每次处理 20-30 篇文献，并要求 AI 跳过已经生成过笔记的条目。这样即使中途失败，也可以继续执行而不会重复写入。 +如果需要和手动 AwesomeGPT 生成结果接近的详细笔记，请明确告诉 AI 使用“满血版”或“深度精读”模式。该模式会读取更多全文内容，并要求模型输出更完整的结构化笔记。 + ## 仓库内容 ```text diff --git a/SKILL.md b/SKILL.md index a080117..3e1f557 100644 --- a/SKILL.md +++ b/SKILL.md @@ -48,6 +48,14 @@ Do not print the config contents. When checking setup, report only whether each py "$env:USERPROFILE\.codex\skills\MYwrite\scripts\generate_zotero_ai_note.py" --vault "C:\Users\qyh15\Documents\Obsidian Vault" --item-key SXAIQUJT --skip-existing ``` +### Deep full-detail item + +Use this when the user asks for `满血版本`, `详细版`, `深度精读`, or says the batch note is not as detailed as manually generated AwesomeGPT notes. Do not use `--skip-existing` when the user wants to regenerate a fuller note; create a new Zotero child note unless the user explicitly asks to delete old notes. + +```powershell +py "$env:USERPROFILE\.codex\skills\MYwrite\scripts\generate_zotero_ai_note.py" --vault "C:\Users\qyh15\Documents\Obsidian Vault" --item-key SXAIQUJT --mode deep --fulltext-chars 80000 --max-tokens 12000 +``` + ### Multiple items ```powershell @@ -92,7 +100,7 @@ Default vault organization: - For live Zotero writes, state that Zotero child notes will be created. - Use `--dry-run` for first-time validation or template changes. -- Use `--fulltext-chars 4000` for cheap, lightweight notes; increase only when the user asks for deeper notes. +- Use `--fulltext-chars 4000` for cheap, lightweight notes; use `--mode deep --fulltext-chars 80000 --max-tokens 12000` when the user asks for a full-detail note. - Do not add visible machine markers to note bodies. Use the Zotero item link for duplicate detection. - Do not delete Zotero duplicate notes unless the user explicitly requests cleanup. - For Obsidian cleanup, delete only files that are empty after removing frontmatter, separators, whitespace, and empty HTML placeholders. diff --git a/scripts/generate_zotero_ai_note.py b/scripts/generate_zotero_ai_note.py index e11670c..62eb183 100644 --- a/scripts/generate_zotero_ai_note.py +++ b/scripts/generate_zotero_ai_note.py @@ -406,7 +406,7 @@ def creators_text(creators: list[dict[str, Any]]) -> str: return "; ".join(names) -def build_prompt(item: dict[str, Any], bibtex: str, fulltext: str, vault: Path) -> str: +def build_prompt(item: dict[str, Any], bibtex: str, fulltext: str, vault: Path, mode: str) -> str: data = item.get("data") or {} prompt_path, research_template_path, review_template_path = template_paths(vault) prompt = prompt_path.read_text(encoding="utf-8") @@ -428,23 +428,37 @@ def build_prompt(item: dict[str, Any], bibtex: str, fulltext: str, vault: Path) "bibtex": bibtex, "indexedFullTextExcerpt": fulltext, } - return "\n\n".join( + instructions = [ + prompt, + "请先判断文献类型：综述型文献或研究型文献。", + "若为研究型文献，请严格填充下面的研究型模板：", + research_template, + "若为综述型文献，请严格填充下面的综述型模板：", + review_template, + "请将模板中的 ${topItem.getField('title')} 替换为真实题名，将 ${topItem.key} 替换为 Zotero key。", + ] + if mode == "deep": + instructions.extend( + [ + "这是满血精读模式。请优先追求完整、具体、可复用的文献阅读笔记，而不是短摘要。", + "每个模板栏目都要充分展开；如果原文提供了材料配比、制备参数、测试条件、器件结构、性能数据、机理解释或对照实验，必须具体写出。", + "核心数据必须尽量保留数值和单位；不要用'显著提高'、'性能优异'这类空泛表达替代具体结果。", + "机制部分要写清因果链条：材料/结构设计如何影响微观结构、界面、导电网络、离子/电子传输、力学响应或器件输出。", + "局限与启发部分要给出可用于后续课题设计的具体启发，而不是泛泛评价。", + "如果全文摘录中缺少某项信息，请明确写'原文摘录中未提供'，不要编造。", + ] + ) + instructions.extend( [ - prompt, - "请先判断文献类型：综述型文献或研究型文献。", - "若为研究型文献，请严格填充下面的研究型模板：", - research_template, - "若为综述型文献，请严格填充下面的综述型模板：", - review_template, - "请将模板中的 ${topItem.getField('title')} 替换为真实题名，将 ${topItem.key} 替换为 Zotero key。", "只输出最终 Markdown 笔记，不要输出解释、判断过程或代码围栏。", "文献材料如下：", json.dumps(source, ensure_ascii=False, indent=2), ] ) + return "\n\n".join(instructions) -def call_llm(prompt: str) -> str: +def call_llm(prompt: str, max_tokens: int | None = None) -> str: api_key = os.environ.get("AWESOMEGPT_API_KEY") base_url = (os.environ.get("AWESOMEGPT_BASE_URL") or "").rstrip("/") model = os.environ.get("AWESOMEGPT_MODEL") @@ -460,6 +474,8 @@ def call_llm(prompt: str) -> str: ], "temperature": 0.3, } + if max_tokens: + payload["max_tokens"] = max_tokens response = http_json( base_url + "/chat/completions", method="POST", @@ -524,6 +540,8 @@ def main() -> None: parser.add_argument("--all", action="store_true", help="Process all top-level Zotero items") parser.add_argument("--limit", type=int, default=1, help="Maximum number of items to process") parser.add_argument("--fulltext-chars", type=int, default=12000) + parser.add_argument("--mode", choices=["quick", "deep"], default="quick", help="quick for batch notes; deep for full-detail notes") + parser.add_argument("--max-tokens", type=int, default=0, help="Optional LLM output token limit; deep mode defaults to 12000") parser.add_argument("--skip-existing", action="store_true", help="Skip items that already have a generated AI child note") parser.add_argument("--dry-run", action="store_true", help="generate but do not write Zotero note") parser.add_argument("--vault", default=str(DEFAULT_VAULT), help="Obsidian vault containing 00 Templater") @@ -557,9 +575,15 @@ def main() -> None: continue print(f"[{index}/{len(items)}] generating {key}: {title}", file=sys.stderr) bibtex = export_bibtex(key) - fulltext = local_fulltext(key, args.fulltext_chars) - prompt = build_prompt(item, bibtex, fulltext, vault) - markdown = call_llm(prompt) + fulltext_chars = args.fulltext_chars + if args.mode == "deep" and fulltext_chars == 12000: + fulltext_chars = 80000 + max_tokens = args.max_tokens + if args.mode == "deep" and not max_tokens: + max_tokens = 12000 + fulltext = local_fulltext(key, fulltext_chars) + prompt = build_prompt(item, bibtex, fulltext, vault, args.mode) + markdown = call_llm(prompt, max_tokens or None) result = create_child_note(user_id, key, markdown, args.dry_run) results.append({"index": index, "itemKey": key, "title": title, "status": "ok", "result": result}) except SystemExit as exc: