sheaf-ai 0.4.0a0__tar.gz

sheaf_ai-0.4.0a0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 zhelunSun
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

sheaf_ai-0.4.0a0/PKG-INFO

@@ -0,0 +1,254 @@
+Metadata-Version: 2.2
+Name: sheaf-ai
+Version: 0.4.0a0
+Summary: Sheaf — Your personal knowledge layer. Paste a link, AI does the rest.
+Author: zhelunSun
+License: MIT
+Project-URL: Homepage, https://github.com/zhelunSun/sheaf-ai
+Project-URL: Repository, https://github.com/zhelunSun/sheaf-ai
+Project-URL: Issues, https://github.com/zhelunSun/sheaf-ai/issues
+Keywords: knowledge-management,ai,mcp,cli,agent
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Information Analysis
+Classifier: Topic :: Internet :: WWW/HTTP :: Indexing/Search
+Classifier: Framework :: Pytest
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: openai>=2.32.0
+Requires-Dist: requests>=2.32.0
+Requires-Dist: beautifulsoup4>=4.12.0
+Requires-Dist: numpy>=1.24.0
+Provides-Extra: browser
+Requires-Dist: playwright>=1.59.0; extra == "browser"
+Provides-Extra: cards
+Requires-Dist: numpy>=1.24.0; extra == "cards"
+Provides-Extra: server
+Requires-Dist: fastapi>=0.100.0; extra == "server"
+Requires-Dist: uvicorn>=0.20.0; extra == "server"
+Requires-Dist: httpx>=0.24.0; extra == "server"
+Provides-Extra: all
+Requires-Dist: playwright>=1.59.0; extra == "all"
+Requires-Dist: numpy>=1.24.0; extra == "all"
+Requires-Dist: fastapi>=0.100.0; extra == "all"
+Requires-Dist: uvicorn>=0.20.0; extra == "all"
+Requires-Dist: httpx>=0.24.0; extra == "all"
+Provides-Extra: dev
+Requires-Dist: pytest>=8.0; extra == "dev"
+Requires-Dist: ruff>=0.4; extra == "dev"
+Requires-Dist: build>=1.0; extra == "dev"
+Requires-Dist: fastapi>=0.100.0; extra == "dev"
+Requires-Dist: uvicorn>=0.20.0; extra == "dev"
+Requires-Dist: httpx>=0.24.0; extra == "dev"
+
+# Sheaf
+
+> **Harvest your knowledge. Bundle it. Share it.**
+
+[![Python 3.10+](https://img.shields.io/badge/python-3.10%2B-blue.svg)](https://www.python.org/downloads/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
+[![Tests](https://img.shields.io/badge/tests-104%20pass-brightgreen)](tests/)
+[![PyPI](https://img.shields.io/pypi/v/sheaf-ai.svg)](https://pypi.org/project/sheaf-ai/)
+[![PyPI - Python Version](https://img.shields.io/pypi/pyversions/sheaf-ai.svg)](https://pypi.org/project/sheaf-ai/)
+
+A **sheaf** is a bundle of grain — the basic unit a farmer brings to market. Sheaf does the same for knowledge: gather what you read, crystallize it into structured bundles, and make it tradable. Your AI agents can search, cite, and reason over everything you've collected.
+
+## Quick Start
+
+```bash
+# Install from PyPI
+pip install sheaf-ai
+
+# Or install from source
+git clone https://github.com/zhelunSun/sheaf-ai.git
+cd sheaf-ai
+pip install -e .
+
+# Set your LLM API key (any OpenAI-compatible endpoint)
+export OPENAI_API_KEY=sk-...
+
+# First-time onboarding (collects 3 sample articles)
+sheaf init
+
+# Collect a link
+sheaf collect https://arxiv.org/abs/2401.00000
+
+# Search your collection
+sheaf search "transformer architecture"
+
+# Crystallize knowledge cards from collected articles
+sheaf crystallize AI
+```
+
+No accounts. No cloud. Your data lives in `./data/` as Markdown + JSON.
+
+## The Problem
+
+You save links every day — articles, repos, papers, tutorials. **95% never get opened again.**
+
+Not because you're lazy. Because bookmarks serve *human reading*, not *agent workflows*. When you ask your coding agent "what did I read about MCP last week?", it has no idea.
+
+Sheaf fixes this. Every link you save becomes a **structured entry** — a single stalk of grain. Crystallize enough of them, and you get a **bundle**: a portable, searchable knowledge pack any agent can consume.
+
+## What It Does
+
+1. **Harvest** — paste a link, Sheaf fetches, classifies, and summarizes it
+2. **Crystallize** — distill 3+ related entries into structured knowledge cards with evidence tracing
+3. **Bundle** — package cards into a portable `.sheaf` unit (coming soon)
+4. **Agent-ready** — built-in MCP server lets any LLM agent query your knowledge base
+
+## Core Commands
+
+```bash
+sheaf collect <url>         # Collect an article, paper, or webpage
+sheaf search <query>        # Full-text search across your collection
+sheaf stats                 # Collection statistics with topic trends
+sheaf crystallize <topic>   # Crystallize knowledge cards from a topic
+sheaf crystallize --list    # List all crystallized cards
+sheaf crystallize --semantic <q>  # Semantic vector search across cards
+sheaf tags                  # Tag statistics
+sheaf weekly                # Weekly summary report
+sheaf insights              # Cross-topic association discovery
+sheaf urgent                # Show entries with upcoming deadlines
+sheaf mcp                   # Start MCP server (stdio transport)
+sheaf init                  # First-time onboarding with demo
+```
+
+## Crystallize: Your Second Brain
+
+This is Sheaf's killer feature. Instead of leaving your bookmarks to rot, `sheaf crystallize` synthesizes insights across multiple entries:
+
+```bash
+$ sheaf crystallize AI
+Crystallizing 'AI'...
+✨ 5 knowledge cards crystallized:
+  📌 RAG faces retrieval relevance challenges (90%)
+     RAG systems heavily depend on retrieval quality; errors degrade LLM output reliability.
+  📌 CRAG framework improves RAG robustness (95%)
+     CRAG introduces a retrieval evaluator, web search augmentation, and document decomposition.
+  📌 Retrieval granularity significantly impacts performance (90%)
+     Finer-grained units like propositions outperform traditional passage-level retrieval.
+```
+
+Each card includes:
+- **Confidence score** (0-100%)
+- **Evidence tracing** — which source entries contributed
+- **Topic provenance** — what topic this card belongs to
+- **Tags** — for filtering and cross-referencing
+
+Use `sheaf crystallize --semantic "query"` for vector-based semantic search across all your cards.
+
+## MCP Server
+
+Sheaf ships with a built-in [Model Context Protocol](https://modelcontextprotocol.io/) server. Any MCP-compatible agent can query your knowledge base:
+
+```bash
+sheaf mcp
+```
+
+**Available tools (9 total):**
+
+| Tool | Description |
+|------|-------------|
+| `sheaf_search` | Full-text search across all entries |
+| `sheaf_list` | List recent entries with filtering |
+| `sheaf_get` | Get full entry details by ID |
+| `sheaf_urgent` | Find time-sensitive entries (deadlines, CFPs) |
+| `sheaf_collect` | Add a new URL to your collection |
+| `sheaf_correct` | Correct a classification error |
+| `sheaf_crystallize` | Crystallize knowledge cards from a topic |
+| `sheaf_list_cards` | List crystallized cards (optional topic filter) |
+| `sheaf_get_card` | Get full card details by ID |
+
+## What You Can Collect
+
+Sheaf handles more than just web articles:
+
+| Input | Example | What Sheaf does |
+|-------|---------|-----------------|
+| **Web articles** | `sheaf collect https://arxiv.org/abs/2401.00000` | Fetches full text, extracts title/author/abstract, classifies topic |
+| **AI chat shares** | `sheaf collect https://chatgpt.com/share/...` | Extracts the Q&A conversation, structures it as reusable knowledge |
+| **WeChat / Zhihu posts** | `sheaf collect https://mp.weixin.qq.com/s/...` | Handles paywalls and dynamic rendering via Playwright fallback |
+| **Pasted text** | `sheaf collect --text "Key insight..."` | Wraps freeform text into a structured entry with auto-classification |
+
+Under the hood, every input goes through the same pipeline: **fetch → classify → summarize → store**. The output is always a structured entry your agents can search and cite.
+
+## Architecture
+
+```
+URL → fetch → classify → summarize → store → query
+         ↓          ↓          ↓         ↓
+    3-strategy   LLM tags   summary   JSONL + MD
+    fallback     + topics   + deadline  index
+
+              ↓
+         crystallize → KnowledgeCard → EmbeddingEngine
+              ↓              ↓
+          CLI/MCP       semantic search
+```
+
+| Module | Purpose |
+|--------|---------|
+| `sheaf_ai/` | Core — pipeline, storage, search, CLI, MCP server, crystallize engine |
+| `sheaf_cards/` | Knowledge card engine — base types, embeddings, generation |
+| `prompts/` | LLM prompt templates (classify, summarize, crystallize) |
+| `data/` | Local knowledge base (JSONL + Markdown, gitignored) |
+
+## Privacy & Local-First
+
+**Your data never leaves your machine unless you choose to.**
+
+- All content stored locally in `./data/` (configurable via `SHEAF_DATA_DIR`)
+- LLM calls go to **your** chosen API provider
+- No telemetry, no analytics, no accounts
+- Markdown + JSONL format — fully portable, zero lock-in
+
+## Configuration
+
+Sheaf works with any OpenAI-compatible API:
+
+```bash
+# OpenAI
+export OPENAI_API_KEY=sk-...
+
+# Or any compatible endpoint (Together, Groq, DeepSeek, etc.)
+export OPENAI_API_KEY=sk-...
+export OPENAI_BASE_URL=https://api.together.xyz/v1
+```
+
+Optional: create a `.env` file in your working directory. See [.env.example](.env.example) for all options.
+
+## Requirements
+
+- **Python 3.10+**
+- **An LLM API key** — any OpenAI-compatible endpoint
+- **Playwright Chromium** (optional, for JS-heavy sites): `pip install -e ".[browser]" && playwright install chromium`
+
+## Development
+
+```bash
+git clone https://github.com/zhelunSun/sheaf-ai.git
+cd sheaf-ai
+pip install -e ".[dev]"
+pytest tests/ -v     # 104 tests
+ruff check sheaf_ai/ tests/ sheaf_cards/
+```
+
+## Alpha Status
+
+Sheaf is in early alpha. The core collect → search → crystallize → MCP pipeline works and is tested with 104 tests. We're validating with real users before beta.
+
+**Try it:** save 20+ links, run `sheaf crystallize <topic>`, then ask your agent to find them. If it works for you, open an issue or discussion to tell us what you'd change.
+
+## License
+
+[MIT](LICENSE)
+
+---
+
+*A **sheaf** is a bundle of harvested grain — the unit a farmer brings to market. In mathematics, a [sheaf](https://en.wikipedia.org/wiki/Sheaf_(mathematics)) attaches local data to open sets and glues them into a global picture. Sheaf the tool does both: gather scattered knowledge into coherent bundles, ready for your agents to consume or for you to share.*

sheaf_ai-0.4.0a0/README.md

@@ -0,0 +1,206 @@
+# Sheaf
+
+> **Harvest your knowledge. Bundle it. Share it.**
+
+[![Python 3.10+](https://img.shields.io/badge/python-3.10%2B-blue.svg)](https://www.python.org/downloads/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
+[![Tests](https://img.shields.io/badge/tests-104%20pass-brightgreen)](tests/)
+[![PyPI](https://img.shields.io/pypi/v/sheaf-ai.svg)](https://pypi.org/project/sheaf-ai/)
+[![PyPI - Python Version](https://img.shields.io/pypi/pyversions/sheaf-ai.svg)](https://pypi.org/project/sheaf-ai/)
+
+A **sheaf** is a bundle of grain — the basic unit a farmer brings to market. Sheaf does the same for knowledge: gather what you read, crystallize it into structured bundles, and make it tradable. Your AI agents can search, cite, and reason over everything you've collected.
+
+## Quick Start
+
+```bash
+# Install from PyPI
+pip install sheaf-ai
+
+# Or install from source
+git clone https://github.com/zhelunSun/sheaf-ai.git
+cd sheaf-ai
+pip install -e .
+
+# Set your LLM API key (any OpenAI-compatible endpoint)
+export OPENAI_API_KEY=sk-...
+
+# First-time onboarding (collects 3 sample articles)
+sheaf init
+
+# Collect a link
+sheaf collect https://arxiv.org/abs/2401.00000
+
+# Search your collection
+sheaf search "transformer architecture"
+
+# Crystallize knowledge cards from collected articles
+sheaf crystallize AI
+```
+
+No accounts. No cloud. Your data lives in `./data/` as Markdown + JSON.
+
+## The Problem
+
+You save links every day — articles, repos, papers, tutorials. **95% never get opened again.**
+
+Not because you're lazy. Because bookmarks serve *human reading*, not *agent workflows*. When you ask your coding agent "what did I read about MCP last week?", it has no idea.
+
+Sheaf fixes this. Every link you save becomes a **structured entry** — a single stalk of grain. Crystallize enough of them, and you get a **bundle**: a portable, searchable knowledge pack any agent can consume.
+
+## What It Does
+
+1. **Harvest** — paste a link, Sheaf fetches, classifies, and summarizes it
+2. **Crystallize** — distill 3+ related entries into structured knowledge cards with evidence tracing
+3. **Bundle** — package cards into a portable `.sheaf` unit (coming soon)
+4. **Agent-ready** — built-in MCP server lets any LLM agent query your knowledge base
+
+## Core Commands
+
+```bash
+sheaf collect <url>         # Collect an article, paper, or webpage
+sheaf search <query>        # Full-text search across your collection
+sheaf stats                 # Collection statistics with topic trends
+sheaf crystallize <topic>   # Crystallize knowledge cards from a topic
+sheaf crystallize --list    # List all crystallized cards
+sheaf crystallize --semantic <q>  # Semantic vector search across cards
+sheaf tags                  # Tag statistics
+sheaf weekly                # Weekly summary report
+sheaf insights              # Cross-topic association discovery
+sheaf urgent                # Show entries with upcoming deadlines
+sheaf mcp                   # Start MCP server (stdio transport)
+sheaf init                  # First-time onboarding with demo
+```
+
+## Crystallize: Your Second Brain
+
+This is Sheaf's killer feature. Instead of leaving your bookmarks to rot, `sheaf crystallize` synthesizes insights across multiple entries:
+
+```bash
+$ sheaf crystallize AI
+Crystallizing 'AI'...
+✨ 5 knowledge cards crystallized:
+  📌 RAG faces retrieval relevance challenges (90%)
+     RAG systems heavily depend on retrieval quality; errors degrade LLM output reliability.
+  📌 CRAG framework improves RAG robustness (95%)
+     CRAG introduces a retrieval evaluator, web search augmentation, and document decomposition.
+  📌 Retrieval granularity significantly impacts performance (90%)
+     Finer-grained units like propositions outperform traditional passage-level retrieval.
+```
+
+Each card includes:
+- **Confidence score** (0-100%)
+- **Evidence tracing** — which source entries contributed
+- **Topic provenance** — what topic this card belongs to
+- **Tags** — for filtering and cross-referencing
+
+Use `sheaf crystallize --semantic "query"` for vector-based semantic search across all your cards.
+
+## MCP Server
+
+Sheaf ships with a built-in [Model Context Protocol](https://modelcontextprotocol.io/) server. Any MCP-compatible agent can query your knowledge base:
+
+```bash
+sheaf mcp
+```
+
+**Available tools (9 total):**
+
+| Tool | Description |
+|------|-------------|
+| `sheaf_search` | Full-text search across all entries |
+| `sheaf_list` | List recent entries with filtering |
+| `sheaf_get` | Get full entry details by ID |
+| `sheaf_urgent` | Find time-sensitive entries (deadlines, CFPs) |
+| `sheaf_collect` | Add a new URL to your collection |
+| `sheaf_correct` | Correct a classification error |
+| `sheaf_crystallize` | Crystallize knowledge cards from a topic |
+| `sheaf_list_cards` | List crystallized cards (optional topic filter) |
+| `sheaf_get_card` | Get full card details by ID |
+
+## What You Can Collect
+
+Sheaf handles more than just web articles:
+
+| Input | Example | What Sheaf does |
+|-------|---------|-----------------|
+| **Web articles** | `sheaf collect https://arxiv.org/abs/2401.00000` | Fetches full text, extracts title/author/abstract, classifies topic |
+| **AI chat shares** | `sheaf collect https://chatgpt.com/share/...` | Extracts the Q&A conversation, structures it as reusable knowledge |
+| **WeChat / Zhihu posts** | `sheaf collect https://mp.weixin.qq.com/s/...` | Handles paywalls and dynamic rendering via Playwright fallback |
+| **Pasted text** | `sheaf collect --text "Key insight..."` | Wraps freeform text into a structured entry with auto-classification |
+
+Under the hood, every input goes through the same pipeline: **fetch → classify → summarize → store**. The output is always a structured entry your agents can search and cite.
+
+## Architecture
+
+```
+URL → fetch → classify → summarize → store → query
+         ↓          ↓          ↓         ↓
+    3-strategy   LLM tags   summary   JSONL + MD
+    fallback     + topics   + deadline  index
+
+              ↓
+         crystallize → KnowledgeCard → EmbeddingEngine
+              ↓              ↓
+          CLI/MCP       semantic search
+```
+
+| Module | Purpose |
+|--------|---------|
+| `sheaf_ai/` | Core — pipeline, storage, search, CLI, MCP server, crystallize engine |
+| `sheaf_cards/` | Knowledge card engine — base types, embeddings, generation |
+| `prompts/` | LLM prompt templates (classify, summarize, crystallize) |
+| `data/` | Local knowledge base (JSONL + Markdown, gitignored) |
+
+## Privacy & Local-First
+
+**Your data never leaves your machine unless you choose to.**
+
+- All content stored locally in `./data/` (configurable via `SHEAF_DATA_DIR`)
+- LLM calls go to **your** chosen API provider
+- No telemetry, no analytics, no accounts
+- Markdown + JSONL format — fully portable, zero lock-in
+
+## Configuration
+
+Sheaf works with any OpenAI-compatible API:
+
+```bash
+# OpenAI
+export OPENAI_API_KEY=sk-...
+
+# Or any compatible endpoint (Together, Groq, DeepSeek, etc.)
+export OPENAI_API_KEY=sk-...
+export OPENAI_BASE_URL=https://api.together.xyz/v1
+```
+
+Optional: create a `.env` file in your working directory. See [.env.example](.env.example) for all options.
+
+## Requirements
+
+- **Python 3.10+**
+- **An LLM API key** — any OpenAI-compatible endpoint
+- **Playwright Chromium** (optional, for JS-heavy sites): `pip install -e ".[browser]" && playwright install chromium`
+
+## Development
+
+```bash
+git clone https://github.com/zhelunSun/sheaf-ai.git
+cd sheaf-ai
+pip install -e ".[dev]"
+pytest tests/ -v     # 104 tests
+ruff check sheaf_ai/ tests/ sheaf_cards/
+```
+
+## Alpha Status
+
+Sheaf is in early alpha. The core collect → search → crystallize → MCP pipeline works and is tested with 104 tests. We're validating with real users before beta.
+
+**Try it:** save 20+ links, run `sheaf crystallize <topic>`, then ask your agent to find them. If it works for you, open an issue or discussion to tell us what you'd change.
+
+## License
+
+[MIT](LICENSE)
+
+---
+
+*A **sheaf** is a bundle of harvested grain — the unit a farmer brings to market. In mathematics, a [sheaf](https://en.wikipedia.org/wiki/Sheaf_(mathematics)) attaches local data to open sets and glues them into a global picture. Sheaf the tool does both: gather scattered knowledge into coherent bundles, ready for your agents to consume or for you to share.*

sheaf_ai-0.4.0a0/prompts/classify.md

@@ -0,0 +1,53 @@
+## Sheaf — 动态分类 Prompt
+
+给定一篇网页文章（标题 + 正文），提取文章的主题和标签。
+
+**核心原则：不做硬分类，让内容自然归类。**
+
+### 输出格式 (JSON)
+
+```json
+{
+  "topics": [
+    {"name": "主题名称", "confidence": 0.9},
+    {"name": "次要主题", "confidence": 0.4}
+  ],
+  "tags": ["标签1", "标签2", "标签3"],
+  "importance": "high | medium | low",
+  "content_type": "news | analysis | research | tutorial | opinion | event | product | reference",
+  "relevance_note": "一句话解释为什么提取这些主题和标签"
+}
+```
+
+### 主题提取规则
+
+1. **topics 是领域维度的归类**，不是固定分类表。LLM 自由提取，例如：
+   - "AI Agent", "Remote Sensing", "Web3", "Climate", "Investing", "International Relations", "Open Source", "LLM", "Computer Vision"...
+   - 一篇文章可以有 1-3 个主题，每个带 confidence（0-1）
+   - confidence > 0.7 的视为主要主题
+
+2. **tags 是更细粒度的关键词**，用于交叉检索。例如：
+   - 文章讨论 GPT-5 发布 → tags: ["GPT-5", "OpenAI", "大模型", "产品发布"]
+   - 文章讨论 RAG 在遥感中的应用 → tags: ["RAG", "remote sensing", "knowledge retrieval", "geospatial"]
+
+3. **content_type 描述文章体裁**：
+   - `news` — 新闻快讯、动态报道
+   - `analysis` — 深度分析、行业洞察
+   - `research` — 学术论文、研究方法
+   - `tutorial` — 教程、实操指南
+   - `opinion` — 观点评论、博客
+   - `event` — 活动预告、会议征稿
+   - `product` — 产品发布、评测
+   - `reference` — 参考资料、工具清单
+   - `ai_conversation` — AI 对话记录（ChatGPT、Claude 等多轮对话归档）
+
+4. **importance 判断**：
+   - `high` — 突破性信息、与用户研究领域直接相关、重大行业变化
+   - `medium` — 值得了解、有参考价值
+   - `low` — 边缘兴趣、信息增量较小
+
+5. **标签提取原则**：
+   - 具体 > 抽象（"RAG" > "AI技术"）
+   - 保留专有名词原文（DeepSeek、OpenAI、Transformer）
+   - 每篇文章 3-8 个标签
+   - 尽量复用已有常见标签（如 "大模型"、"Agent"、"LLM"、"投资"）

sheaf_ai-0.4.0a0/prompts/crystallize.md

@@ -0,0 +1,38 @@
+## Sheaf — Crystallize 知识结晶 Prompt
+
+你是一个知识合成引擎。给定多篇关于同一主题的文章，从中提炼结构化的知识卡片。
+
+**重要：所有文本输出使用中文。** 保留专有名词的英文原文（如模型名、公司名、框架名），但解释和描述用中文。
+
+### 核心原则
+
+1. **跨源合成** — 优先提取多篇来源共同支持的模式和洞察
+2. **证据追溯** — 每个结论必须标注具体来源（用 [Source 0], [Source 1] 格式）
+3. **矛盾标注** — 如果来源之间有矛盾，注明分歧并降低 confidence
+4. **置信度校准** — confidence 应反映证据质量和来源一致性
+
+### 输出格式 (JSON)
+
+```json
+[
+  {
+    "title": "简洁标题（5-15词）",
+    "claim": "核心知识陈述（1-3句话）",
+    "evidence": "具体证据，标注来源 [Source 0]",
+    "tags": ["tag1", "tag2", "tag3"],
+    "confidence": 0.85,
+    "source_indices": [0, 2]
+  }
+]
+```
+
+### 指导原则
+
+- 每个卡片代表一个独立的知识洞察
+- confidence 评分标准：
+  - 0.9+: 多源一致 + 具体数据支撑
+  - 0.7-0.9: 多源支持或单源强证据
+  - 0.5-0.7: 单源支持，需更多验证
+  - <0.5: 推测性结论，源之间存在矛盾
+- 如果来源内容太薄无法提取有意义的模式，返回空数组 []
+- 只输出 JSON，不要解释文字

sheaf_ai-0.4.0a0/prompts/summarize.md

@@ -0,0 +1,50 @@
+## Sheaf — 摘要 Prompt
+
+给定一篇网页文章（标题 + 正文），生成两个层级的摘要。
+
+**重要：所有文本输出必须使用中文。** 保留专有名词的英文原文（如模型名、公司名、框架名），但解释和描述用中文。
+
+### 输出格式 (JSON)
+
+```json
+{
+  "one_liner": "一句话概括核心要点（≤50字）",
+  "structured": {
+    "core_argument": "主要论点或关键洞察（1-3句话）",
+    "key_data": "文章中提到的重要数字、统计数据或具体证据",
+    "relevance_to_user": "为什么这对一位关注 AI 技术趋势的科研或工程人员很重要",
+    "action_items": "可以采取的行动（例如'试用这个框架'、'关注这个趋势'、'在论文中引用'）",
+    "deadline_or_timing": "从文章中提取的时间敏感信息，如果没有则为 null"
+  },
+  "original_title": "文章标题",
+  "source_author": "作者 / 发布方名称（如果能识别的话）"
+}
+```
+
+### 指导原则
+- 简洁。结构化摘要总计控制在 ~200 字以内。
+- 对于 "relevance_to_user"，假设读者是关注 AI 技术趋势的科研或工程人员。
+- 如果文章与用户已知兴趣明显无关，请诚实说明。
+- 保留所有专有名词（人名、公司名、模型名、框架名、论文名）。
+
+### Deadline/Time-sensitive Information Extraction
+
+**CRITICAL**: Pay special attention to any time-sensitive information in the article. This includes:
+
+1. **Conference/Journal deadlines**: CFP deadlines, submission dates, notification dates
+2. **Event dates**: Workshop dates, meetup dates, hackathon dates, forum dates
+3. **Release dates**: Product launches, model releases, version updates
+4. **Registration deadlines**: Sign-up deadlines, early bird deadlines
+5. **Policy/regulation dates**: Compliance deadlines, effective dates
+
+**Extraction format for `deadline_or_timing`**:
+- If there IS time-sensitive info: Write a natural language sentence that INCLUDES the date in ISO format at the end. Examples:
+  - "交叉学科论坛征稿，截止日期为 2026-05-30"
+  - "GPT-5 发布会定于 2026-06-15"
+  - "ACL 2027 投稿截止 2027-01-15"
+- If there is NO time-sensitive info: set to `null`
+
+**Important**:
+- Always extract the SPECIFIC date, not vague references like "next month"
+- If multiple dates exist, list the most important/urgent one
+- Convert relative dates (e.g., "明天", "下周五") to absolute ISO dates based on current context if possible