diff --git a/web/README.md b/web/README.md index e892119..b612484 100644 --- a/web/README.md +++ b/web/README.md @@ -1,11 +1,11 @@ -# Agent Paper Browser +# Agent Knowledge Atlas -本地论文知识库网页，直接读取仓库里的 `data/index.json` 和 `papers/items/*.md`。 +本地 Agent 论文知识地图，直接读取仓库里的 `data/index.json` 和 `papers/items/*.md`。 ## Run ```bash -python3 web/app.py --host 0.0.0.0 --port 18080 +python3 web/app.py --host 100.114.68.27 --port 18080 ``` 默认 Ollama 地址： @@ -25,5 +25,26 @@ http://100.114.68.27:18080 - 默认摘要模型：`ChatGPT-5.6:fast` - 默认翻译模型：`ChatGPT-5.6:light` - 深度分析模型：`ChatGPT-5.6:large`，只在前端手动选择时使用 +- Atlas 解释：`ChatGPT-5.6:fast` +- 主题导读、阅读路径、主题比较：`ChatGPT-5.6:large`，只在前端手动确认后调用 - Ollama 请求单并发执行，并缓存结果到 `web/cache/ai/` - arXiv 摘要缓存到 `web/cache/arxiv/` + +## Knowledge Design + +先由后端从本地语料轻量计算： + +- 主题节点和主题规模 +- 最近三个月热度 +- 主题共现关系 +- 月度趋势 +- 关键论文候选 +- 主题证据列表 + +按需调用 AI 的高维动作： + +- 解释当前 Atlas +- 生成主题导读 +- 生成阅读路径 +- 比较技术路线 +- 单篇论文摘要、翻译、深度研读 diff --git a/web/app.py b/web/app.py index 05bf4fa..6b46d20 100644 --- a/web/app.py +++ b/web/app.py @@ -5,7 +5,6 @@ from __future__ import annotations import argparse import hashlib -import html import json import os import re @@ -15,6 +14,8 @@ import urllib.error import urllib.parse import urllib.request import xml.etree.ElementTree as ET +from collections import Counter, defaultdict +from itertools import combinations from http import HTTPStatus from http.server import SimpleHTTPRequestHandler, ThreadingHTTPServer from pathlib import Path @@ -34,6 +35,10 @@ DEFAULT_MODELS = { "translate": "ChatGPT-5.6:light", "summary": "ChatGPT-5.6:fast", "deep": "ChatGPT-5.6:large", + "atlas": "ChatGPT-5.6:fast", + "topic": "ChatGPT-5.6:large", + "path": "ChatGPT-5.6:large", + "compare": "ChatGPT-5.6:large", } NS = { @@ -44,6 +49,124 @@ NS = { INDEX_LOCK = threading.Lock() OLLAMA_LOCK = threading.Lock() INDEX_CACHE: dict[str, Any] = {"mtime": 0.0, "papers": [], "by_id": {}} +ATLAS_CACHE: dict[str, Any] = {"mtime": 0.0, "atlas": None} + +TOPIC_PROFILES = { + "agent-evaluation": { + "label": "Evaluation", + "stance": "评估已经成为 Agent 研究的中心层：不只是看最终答案，而是看轨迹、工具调用、失败点、成本和安全边界。", + "question": "如何证明一个 Agent 真的能在长程、不确定、带工具的环境里可靠完成任务？", + }, + "tool-use": { + "label": "Tool Use", + "stance": "工具调用正在从 prompt 技巧变成运行时系统问题：schema、权限、依赖、失败恢复和可观测性都变成核心。", + "question": "模型什么时候应该调用工具，如何调用，失败后如何恢复，工具链如何被治理？", + }, + "memory": { + "label": "Memory", + "stance": "Memory 不再只是 RAG 的附属，而是长程 Agent 的状态管理、经验沉淀和安全攻击面。", + "question": "Agent 应该记什么、忘什么、如何写入、如何检索、如何防止记忆污染？", + }, + "rag": { + "label": "Agentic RAG", + "stance": "RAG 正在从相似度检索走向主动检索、证据规划、多跳推理和工具化知识访问。", + "question": "Agent 如何把检索变成可验证、可迭代、可追踪的推理过程？", + }, + "planning": { + "label": "Planning", + "stance": "Planning 的重点从一次性计划转向执行中的重规划、状态校正和长程任务稳定性。", + "question": "长程任务里，计划如何与真实环境反馈、工具失败和上下文限制协同？", + }, + "agent-safety": { + "label": "Safety", + "stance": "Agent safety 已经从拒答问题转向运行时风险：权限、注入、记忆污染、越权动作和治理平面。", + "question": "如何限制 Agent 的行动边界，同时保留足够的自主性和生产力？", + }, + "coding-agent": { + "label": "Coding Agent", + "stance": "Coding Agent 正在从单题修 bug 走向仓库级、长程、交互式的软件工程流程。", + "question": "如何评估和构建能长期维护真实仓库的编码 Agent？", + }, + "computer-use": { + "label": "Computer Use", + "stance": "GUI/browser/OS Agent 正在成为独立层：视觉状态、动作边界、任务进度和安全确认都需要重新建模。", + "question": "Agent 如何可靠理解屏幕状态，并在真实界面中安全行动？", + }, + "multi-agent": { + "label": "Multi-Agent", + "stance": "Multi-agent 正在从角色扮演走向共享记忆、协作协议、路由、治理和组织行为。", + "question": "多个 Agent 如何共享状态、分工、纠错，并避免多数投票和上下文碎片带来的失败？", + }, + "workflow-agent": { + "label": "Workflow", + "stance": "Workflow Agent 把模型能力接进企业流程，核心不在聊天，而在状态、权限、审计和可恢复执行。", + "question": "怎样把 Agent 放进真实业务流程，并让它可观察、可治理、可回滚？", + }, + "reasoning": { + "label": "Reasoning", + "stance": "Reasoning 在 Agent 中必须和工具、记忆、验证器、环境反馈一起看，而不是孤立的思维链。", + "question": "推理如何真正改善行动，而不是只让轨迹看起来更复杂？", + }, + "world-model": { + "label": "World Model", + "stance": "World model 让 Agent 从反应式调用走向对环境、规则和未来状态的内部建模。", + "question": "Agent 需要什么样的世界模型才能更好地预测行动后果？", + }, + "embodied-agent": { + "label": "Embodied", + "stance": "Embodied Agent 把语言、视觉、规划和物理约束连接起来，是 Agent 能力外延的重要方向。", + "question": "语言模型如何在有物理约束的环境中做可执行决策？", + }, +} + +TOPIC_POSITIONS = { + "agent-evaluation": (51, 37), + "tool-use": (37, 52), + "memory": (52, 61), + "rag": (68, 57), + "planning": (48, 48), + "agent-safety": (62, 35), + "coding-agent": (27, 31), + "computer-use": (29, 68), + "multi-agent": (74, 42), + "workflow-agent": (62, 74), + "reasoning": (42, 24), + "world-model": (78, 68), + "embodied-agent": (82, 82), +} + +KNOWLEDGE_ROUTES = [ + { + "id": "evaluation-runtime", + "title": "从答案评测到运行时评估", + "topics": ["agent-evaluation", "tool-use", "agent-safety", "workflow-agent"], + "summary": "Agent 的评估对象从 final answer 扩展到轨迹、工具调用、权限、失败恢复和上线治理。", + }, + { + "id": "memory-state", + "title": "从 RAG 到状态管理", + "topics": ["rag", "memory", "planning", "agent-safety"], + "summary": "长期记忆、主动检索、写入策略和记忆安全正在把 RAG 推成 Agent 的状态层。", + }, + { + "id": "coding-long-horizon", + "title": "从 SWE-Bench 到长程软件工程", + "topics": ["coding-agent", "agent-evaluation", "memory", "tool-use"], + "summary": "编码 Agent 的问题从单 bug 修复扩展到仓库理解、长程维护、交互澄清和 CI 反馈循环。", + }, + { + "id": "computer-use", + "title": "从工具 API 到真实电脑操作", + "topics": ["computer-use", "tool-use", "planning", "agent-safety"], + "summary": "GUI、浏览器和 OS Agent 让状态识别、动作边界、确认机制和视觉记忆成为新的系统层。", + }, + { + "id": "collective-agents", + "title": "从单体智能到协作系统", + "topics": ["multi-agent", "memory", "agent-evaluation", "workflow-agent"], + "summary": "多 Agent 研究正在走向共享记忆、角色分工、路由、群体失败归因和组织行为评估。", + }, +] def ensure_cache_dirs() -> None: @@ -177,6 +300,249 @@ def topic_counts(papers: list[dict[str, Any]]) -> list[dict[str, Any]]: ] +def month_key(value: str) -> str: + if re.match(r"^\d{4}-\d{2}", value or ""): + return value[:7] + return "unknown" + + +def profile_for(topic: str) -> dict[str, str]: + fallback = { + "label": topic.replace("-", " ").title(), + "stance": f"{topic} 是当前语料中自动识别出的主题，需要进一步人工导读。", + "question": "这个主题与 Agent 系统能力的关系是什么？", + } + return {**fallback, **TOPIC_PROFILES.get(topic, {})} + + +def top_papers_for(papers: list[dict[str, Any]], limit: int = 8, recent: bool = False) -> list[dict[str, Any]]: + ranked = list(papers) + if recent: + ranked.sort(key=lambda item: (item["published_at"], item["collection_score"]), reverse=True) + else: + ranked.sort(key=lambda item: (item["collection_score"], item["published_at"]), reverse=True) + return [paper_public(paper) for paper in ranked[:limit]] + + +def build_atlas() -> dict[str, Any]: + papers, _ = load_papers() + mtime = INDEX_PATH.stat().st_mtime + if ATLAS_CACHE["mtime"] == mtime and ATLAS_CACHE["atlas"]: + return ATLAS_CACHE["atlas"] + + topic_to_papers: dict[str, list[dict[str, Any]]] = defaultdict(list) + topic_months: dict[str, Counter] = defaultdict(Counter) + edge_counts: Counter = Counter() + month_counts: Counter = Counter() + + for paper in papers: + month = month_key(paper["published_at"]) + month_counts[month] += 1 + topics = sorted(set(paper["topics"])) + for topic in topics: + topic_to_papers[topic].append(paper) + topic_months[topic][month] += 1 + for left, right in combinations(topics, 2): + edge_counts[(left, right)] += 1 + + months = [month for month in sorted(month_counts) if month != "unknown"] + recent_months = set(months[-3:]) + max_count = max((len(items) for items in topic_to_papers.values()), default=1) + max_recent = max( + (sum(topic_months[topic][month] for month in recent_months) for topic in topic_to_papers), + default=1, + ) + + topics = [] + fallback_index = 0 + for topic, topic_papers in sorted(topic_to_papers.items(), key=lambda item: (-len(item[1]), item[0])): + profile = profile_for(topic) + recent_count = sum(topic_months[topic][month] for month in recent_months) + if topic in TOPIC_POSITIONS: + x, y = TOPIC_POSITIONS[topic] + else: + x = 18 + (fallback_index % 5) * 16 + y = 16 + (fallback_index // 5) * 14 + fallback_index += 1 + topics.append( + { + "id": topic, + "label": profile["label"], + "count": len(topic_papers), + "recent_count": recent_count, + "share": round(len(topic_papers) / max(len(papers), 1), 3), + "heat": round(recent_count / max_recent, 3), + "radius": round(16 + 34 * (len(topic_papers) / max_count), 1), + "x": x, + "y": y, + "stance": profile["stance"], + "question": profile["question"], + "key_papers": top_papers_for(topic_papers, 5), + } + ) + + top_topic_ids = {topic["id"] for topic in topics[:14]} + max_edge_count = max(edge_counts.values(), default=1) + edges = [ + { + "source": left, + "target": right, + "count": count, + "strength": round(count / max_edge_count, 3), + } + for (left, right), count in edge_counts.most_common(80) + if left in top_topic_ids and right in top_topic_ids + ] + + timeline = [] + for month in months: + row = {"month": month, "total": month_counts[month], "topics": {}} + for topic in top_topic_ids: + row["topics"][topic] = topic_months[topic][month] + timeline.append(row) + + atlas = { + "generated_at": time.strftime("%Y-%m-%dT%H:%M:%S%z"), + "paper_count": len(papers), + "topic_count": len(topics), + "months": months, + "recent_months": sorted(recent_months), + "topics": topics, + "edges": edges, + "timeline": timeline, + "routes": KNOWLEDGE_ROUTES, + "key_papers": top_papers_for(papers, 12), + "recent_papers": top_papers_for(papers, 12, recent=True), + "design_policy": { + "precomputed": [ + "主题地图", + "主题规模", + "最近热度", + "主题共现关系", + "月度趋势", + "关键论文候选", + "主题证据列表", + ], + "ai_on_demand": [ + "解释当前地图", + "生成主题导读", + "生成阅读路径", + "比较技术路线", + "单篇论文研读", + ], + }, + } + ATLAS_CACHE.update({"mtime": mtime, "atlas": atlas}) + return atlas + + +def topic_focus(topic: str) -> dict[str, Any]: + atlas = build_atlas() + papers, _ = load_papers() + topic_papers = [paper for paper in papers if topic in paper["topics"]] + if not topic_papers: + raise KeyError(topic) + + months = {row["month"]: row["topics"].get(topic, 0) for row in atlas["timeline"]} + neighbor_counts: Counter = Counter() + for paper in topic_papers: + for other in paper["topics"]: + if other != topic: + neighbor_counts[other] += 1 + + profile = profile_for(topic) + routes = [route for route in KNOWLEDGE_ROUTES if topic in route["topics"]] + focus = { + "topic": topic, + "label": profile["label"], + "stance": profile["stance"], + "question": profile["question"], + "count": len(topic_papers), + "months": months, + "neighbors": [ + { + "topic": name, + "label": profile_for(name)["label"], + "count": count, + "stance": profile_for(name)["stance"], + } + for name, count in neighbor_counts.most_common(10) + ], + "routes": routes, + "key_papers": top_papers_for(topic_papers, 12), + "recent_papers": top_papers_for(topic_papers, 12, recent=True), + "evidence": top_papers_for(topic_papers, 40), + "reading_lens": [ + "先看主题问题，再看高分论文，不要从论文列表开始迷路。", + "优先对比相邻主题，它们通常代表真实系统里的交叉问题。", + "把论文当作证据：它支持了哪条技术路线，暴露了哪个未解决问题？", + ], + } + return focus + + +def atlas_context(topic: str = "", topic_b: str = "") -> str: + atlas = build_atlas() + compact = { + "paper_count": atlas["paper_count"], + "recent_months": atlas["recent_months"], + "top_topics": [ + { + "topic": item["id"], + "label": item["label"], + "count": item["count"], + "recent_count": item["recent_count"], + "stance": item["stance"], + } + for item in atlas["topics"][:13] + ], + "strong_edges": atlas["edges"][:24], + "routes": atlas["routes"], + "key_papers": [ + { + "title": paper["title"], + "topics": paper["topics"], + "score": paper["collection_score"], + "published_at": paper["published_at"], + } + for paper in atlas["key_papers"][:10] + ], + } + if topic: + compact["topic_focus"] = topic_focus(topic) + if topic_b: + compact["compare_topic"] = topic_focus(topic_b) + return json.dumps(compact, ensure_ascii=False, indent=2)[:12000] + + +def atlas_prompt_for(mode: str, context: str) -> str: + no_think = "不要输出思考过程，不要输出 标签，只输出最终结果。" + if mode == "topic": + return ( + "你是 Agent 研究编辑。请根据给定知识库统计，为当前主题写一份中文导读：\n" + "1. 主题定位\n2. 为什么最近重要\n3. 主要技术路线\n4. 关键争议\n" + "5. 必读论文阅读顺序\n6. 和相邻主题的关系\n7. 可以做的实验。\n" + f"只基于给定数据，不要编造论文细节。{no_think}\n\n{context}" + ) + if mode == "path": + return ( + "你是研究导师。请基于给定知识库，为一个想系统学习 Agent 技术的人生成阅读路径。\n" + "输出 4 个阶段：入门地图、核心系统、风险与评估、专题深入。每阶段给出主题、阅读目标、代表论文标题。\n" + f"保持可执行，避免泛泛而谈。{no_think}\n\n{context}" + ) + if mode == "compare": + return ( + "你是技术路线分析师。请比较给定两个 Agent 主题：共同问题、差异、交叉区域、代表论文、工程启发。\n" + f"只使用给定数据。{no_think}\n\n{context}" + ) + return ( + "你是 Agent 知识库的研究编辑。请解释这张 Agent 论文 Atlas：\n" + "1. 当前版图的主轴是什么\n2. 哪些主题是中心节点\n3. 哪些方向最近升温\n" + "4. 哪些主题之间关系最强\n5. 下一步应该优先精读什么。\n" + f"只基于给定数据。{no_think}\n\n{context}" + ) + + def arxiv_id_from_paper(paper: dict[str, Any]) -> str: url = str(paper.get("url") or "") match = re.search(r"arxiv\.org/(?:abs|html|pdf)/([0-9]{4}\.[0-9]+)", url) @@ -288,6 +654,8 @@ def call_ollama(model: str, prompt: str, mode: str) -> dict[str, Any]: options.update({"num_ctx": 3072, "num_predict": 700, "temperature": 0.1}) elif mode == "deep": options.update({"num_ctx": 8192, "num_predict": 900, "temperature": 0.25}) + elif mode in {"atlas", "topic", "path", "compare"}: + options.update({"num_ctx": 8192, "num_predict": 950, "temperature": 0.22}) payload = { "model": model, @@ -354,6 +722,12 @@ class PaperBrowserHandler(SimpleHTTPRequestHandler): papers, _ = load_papers() write_json(self, {"topics": topic_counts(papers)}) return + if path == "/api/atlas": + write_json(self, build_atlas()) + return + if path == "/api/topic": + self.handle_topic(query) + return if path == "/api/papers": self.handle_paper_search(query) return @@ -371,6 +745,9 @@ class PaperBrowserHandler(SimpleHTTPRequestHandler): if parsed.path == "/api/ai": self.handle_ai() return + if parsed.path == "/api/atlas/ai": + self.handle_atlas_ai() + return write_error(self, HTTPStatus.NOT_FOUND, "not found") def serve_file(self, target: Path, content_type: str) -> None: @@ -479,6 +856,16 @@ class PaperBrowserHandler(SimpleHTTPRequestHandler): text = (ROOT / paper["path"]).read_text(encoding="utf-8") write_json(self, {"paper": paper_public(paper), "markdown": text}) + def handle_topic(self, query: dict[str, list[str]]) -> None: + topic = (query.get("topic") or [""])[0].strip() + if not topic: + write_error(self, HTTPStatus.BAD_REQUEST, "topic is required") + return + try: + write_json(self, topic_focus(topic)) + except KeyError: + write_error(self, HTTPStatus.NOT_FOUND, "topic not found") + def handle_abstract(self, query: dict[str, list[str]]) -> None: paper_id = (query.get("id") or [""])[0] _, by_id = load_papers() @@ -545,6 +932,53 @@ class PaperBrowserHandler(SimpleHTTPRequestHandler): except Exception as exc: # noqa: BLE001 write_error(self, HTTPStatus.INTERNAL_SERVER_ERROR, str(exc)) + def handle_atlas_ai(self) -> None: + try: + payload = read_json_body(self) + mode = str(payload.get("mode") or "atlas") + topic = str(payload.get("topic") or "") + topic_b = str(payload.get("topic_b") or "") + refresh = bool(payload.get("refresh")) + model = str(payload.get("model") or DEFAULT_MODELS.get(mode) or DEFAULT_MODELS["atlas"]) + if mode not in {"atlas", "topic", "path", "compare"}: + write_error(self, HTTPStatus.BAD_REQUEST, "invalid atlas ai mode") + return + if mode == "topic" and not topic: + write_error(self, HTTPStatus.BAD_REQUEST, "topic is required") + return + if mode == "compare" and (not topic or not topic_b): + write_error(self, HTTPStatus.BAD_REQUEST, "topic and topic_b are required") + return + + ensure_cache_dirs() + cache_key = f"atlas:{mode}:{topic}:{topic_b}" + cache_path = ai_cache_path(cache_key, mode, model) + if cache_path.exists() and not refresh: + result = json.loads(cache_path.read_text(encoding="utf-8")) + result["cached"] = True + write_json(self, result) + return + + if not OLLAMA_LOCK.acquire(blocking=False): + write_error(self, HTTPStatus.TOO_MANY_REQUESTS, "ollama is busy; try again later") + return + try: + context = atlas_context(topic, topic_b) + result = call_ollama(model, atlas_prompt_for(mode, context), mode) + result.update( + { + "scope": cache_key, + "cached": False, + "created_at": time.strftime("%Y-%m-%dT%H:%M:%S%z"), + } + ) + cache_path.write_text(json.dumps(result, ensure_ascii=False, indent=2) + "\n", encoding="utf-8") + write_json(self, result) + finally: + OLLAMA_LOCK.release() + except Exception as exc: # noqa: BLE001 + write_error(self, HTTPStatus.INTERNAL_SERVER_ERROR, str(exc)) + def parse_args() -> argparse.Namespace: parser = argparse.ArgumentParser() diff --git a/web/static/app.js b/web/static/app.js index 2cf82b4..f8db2bd 100644 --- a/web/static/app.js +++ b/web/static/app.js @@ -1,14 +1,14 @@ const state = { - papers: [], - total: 0, - offset: 0, - limit: 50, - selectedId: null, + atlas: null, + health: null, + view: "atlas", + activeTopic: "", + topicFocus: null, selectedPaper: null, selectedMarkdown: "", - topic: "", - loading: false, - health: null, + searchItems: [], + searchTotal: 0, + aiBusy: false, }; const $ = (selector) => document.querySelector(selector); @@ -16,28 +16,17 @@ const $ = (selector) => document.querySelector(selector); const elements = { healthLine: $("#healthLine"), searchInput: $("#searchInput"), - yearSelect: $("#yearSelect"), - statusSelect: $("#statusSelect"), - sortSelect: $("#sortSelect"), - clearTopicBtn: $("#clearTopicBtn"), - topicChips: $("#topicChips"), - resultTitle: $("#resultTitle"), - resultCount: $("#resultCount"), - resultsList: $("#resultsList"), - loadMoreBtn: $("#loadMoreBtn"), - detailMetaLine: $("#detailMetaLine"), - detailTitle: $("#detailTitle"), - detailTags: $("#detailTags"), - arxivLink: $("#arxivLink"), - paperPreview: $("#paperPreview"), - abstractBtn: $("#abstractBtn"), - summaryBtn: $("#summaryBtn"), - translateBtn: $("#translateBtn"), - deepBtn: $("#deepBtn"), - modelStatus: $("#modelStatus"), - aiStatus: $("#aiStatus"), - abstractView: $("#abstractView"), - aiView: $("#aiView"), + searchBtn: $("#searchBtn"), + mapTitle: $("#mapTitle"), + atlasSvg: $("#atlasSvg"), + routeList: $("#routeList"), + timeline: $("#timeline"), + explainAtlasBtn: $("#explainAtlasBtn"), + readingPathBtn: $("#readingPathBtn"), + lensEyebrow: $("#lensEyebrow"), + lensTitle: $("#lensTitle"), + lensBody: $("#lensBody"), + closePaperBtn: $("#closePaperBtn"), }; function escapeHtml(value) { @@ -49,7 +38,7 @@ function escapeHtml(value) { .replaceAll("'", "'"); } -function debounce(fn, delay = 250) { +function debounce(fn, delay = 260) { let timer = null; return (...args) => { window.clearTimeout(timer); @@ -72,98 +61,384 @@ async function api(path, options = {}) { return payload; } -function queryParams(extra = {}) { - const params = new URLSearchParams(); - const q = elements.searchInput.value.trim(); - if (q) params.set("q", q); - if (state.topic) params.set("topic", state.topic); - if (elements.yearSelect.value) params.set("year", elements.yearSelect.value); - if (elements.statusSelect.value) params.set("status", elements.statusSelect.value); - if (elements.sortSelect.value) params.set("sort", elements.sortSelect.value); - params.set("limit", String(state.limit)); - params.set("offset", String(state.offset)); - Object.entries(extra).forEach(([key, value]) => params.set(key, String(value))); - return params; +function topicLabel(topic) { + const found = state.atlas?.topics?.find((item) => item.id === topic); + return found?.label || topic.replaceAll("-", " "); +} + +function colorFor(index) { + const colors = ["#0d766f", "#ba5644", "#aa7419", "#527b46", "#665d9f", "#236092", "#8a5a25"]; + return colors[index % colors.length]; } async function loadHealth() { try { const health = await api("/api/health"); state.health = health; - elements.healthLine.textContent = `${health.paper_count} papers`; - elements.modelStatus.textContent = health.ollama_ok - ? `Ollama 已连接 · summary=${health.default_models.summary}` - : "Ollama 未连接"; + elements.healthLine.textContent = `${health.paper_count} papers · ${health.ollama_ok ? "Ollama online" : "Ollama offline"}`; } catch (error) { - elements.healthLine.textContent = "服务状态未知"; - elements.modelStatus.textContent = error.message; + elements.healthLine.textContent = error.message; } } -async function loadTopics() { - const payload = await api("/api/topics"); - elements.topicChips.innerHTML = payload.topics - .slice(0, 24) - .map((item) => { - const active = item.topic === state.topic ? " active" : ""; - return ``; - }) - .join(""); +async function loadAtlas() { + state.atlas = await api("/api/atlas"); + elements.mapTitle.textContent = `${state.atlas.paper_count} 篇论文构成的 Agent 版图`; + renderAtlasMap(); + renderRoutes(); + renderTimeline(); + renderOverview(); } -async function loadPapers({ reset = false } = {}) { - if (state.loading) return; - state.loading = true; - if (reset) { - state.offset = 0; - state.papers = []; - } - elements.resultTitle.textContent = state.topic || "论文库"; - elements.resultCount.textContent = "..."; - try { - const payload = await api(`/api/papers?${queryParams()}`); - state.total = payload.total; - state.papers = reset ? payload.items : [...state.papers, ...payload.items]; - state.offset = state.papers.length; - renderResults(); - if (!state.selectedId && state.papers.length) { - await selectPaper(state.papers[0].id); - } - } catch (error) { - elements.resultsList.innerHTML = `
${escapeHtml(error.message)}
`; - } finally { - state.loading = false; - } -} +function renderAtlasMap() { + const topics = state.atlas.topics.filter((topic) => topic.count >= 30).slice(0, 13); + const topicMap = new Map(topics.map((topic, index) => [topic.id, { ...topic, index }])); + const maxCount = Math.max(...topics.map((topic) => topic.count), 1); + const isMobile = window.matchMedia("(max-width: 760px)").matches; + elements.atlasSvg.setAttribute("viewBox", isMobile ? "16 -4 78 86" : "0 0 170 100"); + const point = (topic) => ({ + x: isMobile ? topic.x : 14 + topic.x * 1.58, + y: -18 + topic.y * 0.93, + }); -function renderResults() { - elements.resultCount.textContent = String(state.total); - elements.loadMoreBtn.style.display = state.papers.length < state.total ? "block" : "none"; - if (!state.papers.length) { - elements.resultsList.innerHTML = `
没有匹配结果
`; - return; - } - elements.resultsList.innerHTML = state.papers - .map((paper) => { - const active = paper.id === state.selectedId ? " active" : ""; - const topics = paper.topics - .slice(0, 4) - .map((topic) => `${escapeHtml(topic)}`) - .join(""); + const edges = state.atlas.edges + .filter((edge) => topicMap.has(edge.source) && topicMap.has(edge.target)) + .map((edge) => { + const source = topicMap.get(edge.source); + const target = topicMap.get(edge.target); + const sourcePoint = point(source); + const targetPoint = point(target); return ` - + + ${escapeHtml(topicLabel(edge.source))} ↔ ${escapeHtml(topicLabel(edge.target))}: ${edge.count} + `; }) .join(""); + + const nodes = topics + .map((topic, index) => { + const active = topic.id === state.activeTopic ? " active" : ""; + const r = 3.6 + 5.6 * (topic.count / maxCount); + const ring = r + 1.6 + topic.heat * 1.7; + const color = colorFor(index); + const p = point(topic); + return ` + + + + ${escapeHtml(topic.label)} + ${topic.count} · +${topic.recent_count} + + `; + }) + .join(""); + + elements.atlasSvg.innerHTML = `${edges}${nodes}`; +} + +function renderRoutes() { + elements.routeList.innerHTML = state.atlas.routes + .map( + (route) => ` + + `, + ) + .join(""); +} + +function renderTimeline() { + const rows = state.atlas.timeline.slice(-13); + const maxTotal = Math.max(...rows.map((row) => row.total), 1); + elements.timeline.innerHTML = rows + .map((row) => { + const height = 8 + 70 * (row.total / maxTotal); + return ` +
+
+
${escapeHtml(row.month.slice(5))}
+
+ `; + }) + .join(""); +} + +function statGrid(items) { + return ` +
+ ${items + .map( + (item) => ` +
+ ${escapeHtml(item.value)} + ${escapeHtml(item.label)} +
+ `, + ) + .join("")} +
+ `; +} + +function paperList(items, limit = items.length) { + if (!items?.length) { + return `
暂无证据
`; + } + return ` +
+ ${items + .slice(0, limit) + .map( + (paper) => ` + + `, + ) + .join("")} +
+ `; +} + +function routeButtons(routes) { + return ` +
+ ${routes + .map( + (route) => ` + + `, + ) + .join("")} +
+ `; +} + +function renderOverview(aiHtml = "") { + state.view = "atlas"; + state.selectedPaper = null; + elements.lensEyebrow.textContent = "Atlas"; + elements.lensTitle.textContent = "整体脉络"; + setModeButtons("atlas"); + + const topTopics = state.atlas.topics.slice(0, 7); + elements.lensBody.innerHTML = ` +
+ ${statGrid([ + { value: state.atlas.paper_count, label: "papers" }, + { value: state.atlas.topic_count, label: "topics" }, + { value: state.atlas.recent_months.join(" / "), label: "recent window" }, + ])} +
+ +
+

这版 Atlas 把论文当作证据，把主题、路线和趋势当作入口。当前最强的主轴是 evaluation、tool use、RAG、reasoning、planning、safety、memory。

+

核心问题：Agent 如何从“能回答”走向“能长期、可观察、可治理地行动”？

+
+ +
+

主题入口

+
+ ${topTopics + .map( + (topic) => ` + + `, + ) + .join("")} +
+
+ +
+

路线

+ ${routeButtons(state.atlas.routes)} +
+ +
+

关键论文候选

+ ${paperList(state.atlas.key_papers, 8)} +
+ +
+

AI 研究编辑

+
+ + +
+
${aiHtml}
+
+ `; +} + +async function selectTopic(topic) { + state.activeTopic = topic; + state.selectedPaper = null; + renderAtlasMap(); + const focus = await api(`/api/topic?topic=${encodeURIComponent(topic)}`); + state.topicFocus = focus; + renderTopic(focus); +} + +function renderTopic(focus, aiHtml = "") { + state.view = "atlas"; + elements.lensEyebrow.textContent = "Topic"; + elements.lensTitle.textContent = focus.label; + setModeButtons("atlas"); + + const recentTotal = Object.values(focus.months).slice(-3).reduce((sum, value) => sum + Number(value || 0), 0); + elements.lensBody.innerHTML = ` +
+ ${statGrid([ + { value: focus.count, label: "papers" }, + { value: recentTotal, label: "recent" }, + { value: focus.neighbors.length, label: "neighbors" }, + ])} +
+ +
+

${escapeHtml(focus.stance)}

+

${escapeHtml(focus.question)}

+
+ +
+

相邻主题

+
+ ${focus.neighbors + .slice(0, 8) + .map( + (item) => ` + + `, + ) + .join("")} +
+
+ +
+

所在路线

+ ${routeButtons(focus.routes)} +
+ +
+

AI 研究编辑

+
+ + + ${ + focus.neighbors[0] + ? `` + : "" + } +
+
${aiHtml}
+
+ +
+

关键论文

+ ${paperList(focus.key_papers, 10)} +
+ +
+

最近新增

+ ${paperList(focus.recent_papers, 8)} +
+ `; +} + +function renderRoute(routeId) { + const route = state.atlas.routes.find((item) => item.id === routeId); + if (!route) return; + state.selectedPaper = null; + elements.lensEyebrow.textContent = "Route"; + elements.lensTitle.textContent = route.title; + elements.lensBody.innerHTML = ` +
+

${escapeHtml(route.summary)}

+
+
+

主题链路

+
+ ${route.topics + .map((topic) => ``) + .join("")} +
+
+
+

路线证据

+ ${paperList( + state.atlas.key_papers.filter((paper) => paper.topics.some((topic) => route.topics.includes(topic))), + 12, + )} +
+ `; +} + +async function runSearch() { + const q = elements.searchInput.value.trim(); + if (!q) { + renderOverview(); + return; + } + state.view = "search"; + setModeButtons("search"); + elements.lensEyebrow.textContent = "Search"; + elements.lensTitle.textContent = q; + elements.lensBody.innerHTML = `
searching
`; + const params = new URLSearchParams({ q, limit: "60", sort: "score" }); + if (state.activeTopic) params.set("topic", state.activeTopic); + const payload = await api(`/api/papers?${params.toString()}`); + state.searchItems = payload.items; + state.searchTotal = payload.total; + elements.lensBody.innerHTML = ` +
+ ${statGrid([ + { value: payload.total, label: "matches" }, + { value: state.activeTopic ? topicLabel(state.activeTopic) : "all", label: "scope" }, + { value: "score", label: "sort" }, + ])} +
+
+ ${paperList(payload.items, payload.items.length)} +
+ `; +} + +function renderReading() { + state.view = "reading"; + setModeButtons("reading"); + elements.lensEyebrow.textContent = "Reading"; + elements.lensTitle.textContent = "阅读流"; + elements.lensBody.innerHTML = ` +
+

先从路线理解问题，再进入主题，最后把论文作为证据读。当前版本把高分论文和最近新增放在这里，后续可以接入个人阅读队列。

+
+
+

高优先级

+ ${paperList(state.atlas.key_papers, 12)} +
+
+

最近新增

+ ${paperList(state.atlas.recent_papers, 12)} +
+ `; } function stripFrontmatter(markdown) { @@ -172,196 +447,234 @@ function stripFrontmatter(markdown) { return match ? text.slice(match[0].length) : text; } -function inlineMarkdown(text) { - let value = escapeHtml(text); - value = value.replace(/\[([^\]]+)\]\((https?:\/\/[^)]+)\)/g, '$1'); - value = value.replace(/`([^`]+)`/g, "$1"); - return value; -} - function renderMarkdown(markdown) { - const lines = stripFrontmatter(markdown).split(/\r?\n/); - const output = []; - let inList = false; - const closeList = () => { - if (inList) { - output.push(""); - inList = false; - } - }; - - for (const raw of lines) { - const line = raw.trimEnd(); - if (!line.trim()) { - closeList(); - continue; - } - if (line.startsWith("### ")) { - closeList(); - output.push(`

${inlineMarkdown(line.slice(4))}

`); - } else if (line.startsWith("## ")) { - closeList(); - output.push(`

${inlineMarkdown(line.slice(3))}

`); - } else if (line.startsWith("# ")) { - closeList(); - output.push(`

${inlineMarkdown(line.slice(2))}

`); - } else if (line.startsWith("- ")) { - if (!inList) { - output.push("