pagesift_engine/engine/
mod.rs

1//! 抓取引擎：把网络层、礼貌策略、正文提取、PDF 和浏览器渲染串成一次完整的抓取。
2//!
3//! 流程：robots 模式 → HTTP 抓取（策略层负责冷却、robots、限速）→ 按内容类型分流（PDF / HTML）
4//! → `engine: auto` 时正文太少就升级到浏览器渲染 → 组装成 Firecrawl 格式的文档。
5
6use std::collections::BTreeMap;
7use std::sync::Arc;
8use std::time::{Duration, Instant};
9
10use pagesift_api::types::{Document, Metadata, PagesiftMeta, ScrapeOptions};
11use pagesift_api::{ApiError, BoxFuture, Engine, ErrorCode};
12use pagesift_cdp::{
13    AllowListGuard, BlockResources, BrowserPool, NetworkAllowList, PoolConfig, PrivateNetworkGuard,
14    RenderOptions, RenderedPage, RequestGuard, SubRequest, Verdict,
15};
16use pagesift_net::{FetchRequest, Fetcher, FetcherConfig, Url};
17use pagesift_policy::{Policy, PolicyConfig, RequestKind, RobotsMode};
18use tokio::sync::OnceCell;
19
20pub mod agent;
21mod autoschema;
22mod cache;
23mod cpu;
24mod csv_text;
25mod egress_route;
26mod extract;
27mod feed;
28mod fetch;
29mod frames;
30mod hooked;
31mod http_extract;
32mod image_file;
33mod images;
34mod interact;
35mod office;
36mod output;
37mod parse;
38mod pdf;
39pub mod planner;
40mod render;
41mod render_stats;
42mod scraper;
43mod speculate;
44pub use autoschema::ActionSecret;
45pub use hooked::ScraperBuilder;
46pub use interact::InteractLimits;
47mod similar;
48mod structured;
49use cpu::{CONTENT_WORKSPACE, cpu_task};
50pub use cpu::{set_cpu_capacity, set_cpu_groups};
51use extract::*;
52use frames::*;
53use http_extract::*;
54use office::*;
55pub use office::{office_markdown, office_markdown_from};
56pub use output::*;
57use render::*;
58
59/// F-1.9：HTTP 抓到的正文少于这么多（加权）字符时，`engine: auto` 升级到浏览器渲染。
60/// 中日韩文字一个字按 [`CJK_WEIGHT`] 个字符算（信息量和几个拉丁字母相当），见 [`text_weight`]。
61pub const AUTO_RENDER_MIN_CHARS: usize = 1200;
62/// 已经认出是一篇文章（有发布时间或 `og:type=article`）时，正文到这么多加权字符就不再渲染：
63/// 短新闻 HTTP 已经拿全了。AiSeek 的生产数据里，63% 的渲染拿到的文字和 HTTP 一样多，
64/// 多半就是这类短文章。
65pub const AUTO_RENDER_ARTICLE_MIN_CHARS: usize = 300;
66pub const CJK_WEIGHT: usize = 3;
67
68/// 正文的加权长度：空白不算，中日韩文字按 [`CJK_WEIGHT`] 算。
69pub fn text_weight(s: &str) -> usize {
70    s.chars()
71        .filter(|c| !c.is_whitespace())
72        .map(|c| if is_cjk(c) { CJK_WEIGHT } else { 1 })
73        .sum()
74}
75
76fn is_cjk(c: char) -> bool {
77    matches!(c as u32,
78        0x3040..=0x30FF   // 平假名、片假名
79        | 0x3400..=0x4DBF // CJK 扩展 A
80        | 0x4E00..=0x9FFF // CJK 统一汉字
81        | 0xAC00..=0xD7AF // 谚文音节
82        | 0xF900..=0xFAFF // CJK 兼容汉字
83        | 0x20000..=0x2FA1F)
84}
85
86/// `engine: auto` 时 HTTP 的结果是否还要用浏览器渲染。
87pub fn needs_render(doc: &pagesift_content::Document) -> bool {
88    let w = text_weight(&doc.markdown);
89    if w >= AUTO_RENDER_MIN_CHARS {
90        return false;
91    }
92    let m = &doc.metadata;
93    let article = m.published_time.is_some()
94        || m.extra
95            .iter()
96            .any(|(k, v)| k == "og:type" && v.eq_ignore_ascii_case("article"));
97    !(article && w >= AUTO_RENDER_ARTICLE_MIN_CHARS)
98}
99/// 浏览器渲染出的正文少于这么多字符时，换下一个内核再试（Lightpanda 的脚本支持不全，
100/// 单页应用可能渲染不出内容）。最后一个内核的结果照常返回。
101pub const RENDER_FALLBACK_MIN_CHARS: usize = 200;
102
103/// 内核链里一个内核的启动检查结果（[`Scraper::probe_engines`]）。
104#[derive(Debug, Clone)]
105pub struct EngineProbe {
106    pub engine: &'static str,
107    /// 可执行文件路径或远程地址。
108    pub location: String,
109    pub result: Result<(), String>,
110}
111
112#[derive(Clone, Default)]
113pub struct ScraperConfig {
114    pub fetcher: FetcherConfig,
115    pub policy: PolicyConfig,
116    pub browser: PoolConfig,
117    /// 抓取结果缓存（`maxAge`、`minAge`、`storeInCache`）；不设时不缓存。
118    pub store: Option<pagesift_store::Store>,
119    /// `summary`、`json` 格式和 `/v2/extract` 用的 OpenAI 兼容服务。
120    pub llm: Option<pagesift_llm::LlmConfig>,
121    /// `/v2/search` 用的搜索后端。
122    pub search: Option<Arc<dyn pagesift_search::SearchBackend>>,
123    /// 命名出口（名字，代理地址）：请求的 `egress` 选其中一个，其余抓取设置和默认出口相同；
124    /// 两个及以上时，不指定 `egress` 的请求由出口池选。
125    pub egress: Vec<(String, String)>,
126    /// 命名出口的类别（没列出的是 `direct`）。代理地址为空的 `direct` 出口与默认出口相同。
127    pub egress_classes: std::collections::HashMap<String, pagesift_policy::EgressClass>,
128    /// 审计日志：每次抓取一行元数据，可疑结果留 HTML 快照（见 [`crate::audit`]）。
129    pub audit: Option<Arc<crate::audit::Audit>>,
130    /// 扩展点（`pagesift_hooks`）：代码注册的钩子和 `[[rules]]` 编译出的规则。空时不经过任何钩子。
131    pub hooks: pagesift_hooks::Hooks,
132    /// `[secrets]`：act 动作里 `{{secret:NAME}}` 在本地替换成的值（不经过模型）。
133    pub secrets: std::collections::BTreeMap<String, ActionSecret>,
134    /// 扫描页 OCR（`[ocr] enabled`）；不设时扫描页只有说明，并报 `scanned_pages_without_ocr`。
135    pub ocr: Option<Arc<dyn pagesift_pdf::PageOcr>>,
136    /// `/interact` 保留会话的时长上限（默认 Firecrawl 文档的值）。
137    pub interact: InteractLimits,
138}
139
140/// 创建 [`Scraper`] 失败：网络层或 LLM 客户端的配置不对。
141#[derive(Debug)]
142pub enum ScraperError {
143    Fetcher(pagesift_net::Error),
144    Llm(pagesift_llm::LlmError),
145}
146
147impl std::fmt::Display for ScraperError {
148    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
149        match self {
150            Self::Fetcher(e) => write!(f, "fetcher configuration: {e}"),
151            Self::Llm(e) => write!(f, "LLM configuration: {e}"),
152        }
153    }
154}
155
156impl std::error::Error for ScraperError {
157    fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
158        match self {
159            Self::Fetcher(e) => Some(e),
160            Self::Llm(e) => Some(e),
161        }
162    }
163}
164
165pub struct Scraper {
166    policy: Arc<Policy>,
167    store: Option<pagesift_store::Store>,
168    llm: Option<pagesift_llm::ChatClient>,
169    search: Option<Arc<dyn pagesift_search::SearchBackend>>,
170    ocr: Option<Arc<dyn pagesift_pdf::PageOcr>>,
171    /// 网络层放开了内网（`allowPrivateNetwork`）时，浏览器的子请求也一样放开。
172    browser_private_network: bool,
173    /// 浏览器子请求的细粒度内网放行名单（allow_hosts / allow_ips / allow_cidrs）。
174    browser_allow: NetworkAllowList,
175    /// 按优先级排列的内核（见 [`pagesift_cdp::engine_chain`]）。每个池第一次需要时才创建；
176    /// 创建失败时记住结果，不反复尝试。
177    browsers: Vec<(PoolConfig, OnceCell<Option<Arc<BrowserPool>>>)>,
178    /// 每个内核最近一次启动失败的时间：之后 [`LAUNCH_RETRY`] 内跳过它，不让每个请求都去试一次。
179    launch_failed: Vec<std::sync::Mutex<Option<Instant>>>,
180    /// 每个内核最近一次启动失败的原因，用来在“浏览器不可用”时告诉调用方真因（root、沙箱等）。
181    launch_notes: Vec<std::sync::Mutex<Option<String>>>,
182    /// 有内核找到了浏览器、但创建或启动失败（不是"没找到浏览器"）。
183    launch_broken: std::sync::atomic::AtomicBool,
184    /// 爬取里同时进行的抓取数，等于全局并发。排队的页面只占一个很小的等待 future，
185    /// 拿到名额才创建完整的抓取状态机（十几 KB）：同时爬上千个站时，排队的页面不再各占一份。
186    /// 上限跟着限速器的全局在途数目标走（USL 控制器，见 `pagesift_policy::RateLimiter`）。
187    crawl_slots: pagesift_policy::AdaptiveConcurrencyGate,
188    learner: crate::learn::HostLearner,
189    /// 要不要开浏览器、渲染结果怎么和 HTTP 结果合并。
190    planner: planner::RenderPlanner,
191    /// 自动升级的在线耗时分布：预算、导航耐心、早停时机都从这里来。
192    render_stats: render_stats::RenderStats,
193    audit: Option<Arc<crate::audit::Audit>>,
194    hooks: pagesift_hooks::Hooks,
195    secrets: std::collections::BTreeMap<String, ActionSecret>,
196    /// 多服务器网关：爬取里的页面交给它（工人执行，见 `pagesift` 的 `cluster` 模块）。单主机不设，路径不变。
197    pages: std::sync::OnceLock<Arc<dyn crate::PageRunner>>,
198    /// `/interact` 的会话表。
199    sessions: Arc<interact::Sessions>,
200}
201
202/// 文档从哪来：抓一个网址，或解析调用方上传的文件（`/v2/parse`）。
203pub(super) enum Source<'a> {
204    Url(&'a str, RobotsMode),
205    Upload(pagesift_api::types::Upload),
206}
207
208#[derive(Debug, Clone, Copy, PartialEq, Eq)]
209enum EngineChoice {
210    Http,
211    Browser,
212    Auto,
213}
214
215impl Engine for Scraper {
216    fn scrape<'a>(
217        &'a self,
218        url: &'a str,
219        opts: &'a ScrapeOptions,
220    ) -> BoxFuture<'a, Result<Document, ApiError>> {
221        Box::pin(Scraper::scrape(self, url, opts))
222    }
223
224    fn parse<'a>(
225        &'a self,
226        file: pagesift_api::types::Upload,
227        opts: &'a ScrapeOptions,
228    ) -> BoxFuture<'a, Result<Document, ApiError>> {
229        Box::pin(Scraper::parse(self, file, opts))
230    }
231
232    fn map<'a>(
233        &'a self,
234        req: &'a pagesift_api::types::MapRequest,
235    ) -> BoxFuture<'a, Result<Vec<pagesift_api::types::MapLink>, ApiError>> {
236        Box::pin(Scraper::map(self, req))
237    }
238
239    fn crawl<'a>(
240        &'a self,
241        req: &'a pagesift_api::types::CrawlRequest,
242        sink: Arc<dyn pagesift_api::CrawlSink>,
243    ) -> BoxFuture<'a, Result<(), ApiError>> {
244        Box::pin(Scraper::crawl(self, req, sink))
245    }
246
247    fn search<'a>(
248        &'a self,
249        req: &'a pagesift_api::types::SearchRequest,
250    ) -> BoxFuture<'a, Result<serde_json::Value, ApiError>> {
251        Box::pin(Scraper::search(self, req))
252    }
253
254    fn extract<'a>(
255        &'a self,
256        req: &'a pagesift_api::types::ExtractRequest,
257    ) -> BoxFuture<'a, Result<serde_json::Value, ApiError>> {
258        Box::pin(Scraper::extract(self, req))
259    }
260
261    fn supports_interact(&self) -> bool {
262        self.interact_available()
263    }
264
265    /// 动作只在 Chrome 上验证过，和 interact 同一个条件。
266    fn supports_actions(&self) -> bool {
267        self.interact_available()
268    }
269
270    fn actions_unavailable(&self) -> &'static str {
271        "actions need Chrome: no Chrome or Chromium was found on this machine (run `pagesift doctor`)"
272    }
273
274    fn scrape_session<'a>(
275        &'a self,
276        url: &'a str,
277        opts: &'a ScrapeOptions,
278        keep: bool,
279        owner: Option<&'a str>,
280    ) -> BoxFuture<'a, Result<Document, ApiError>> {
281        Box::pin(Scraper::scrape_session(self, url, opts, keep, owner))
282    }
283
284    fn observe<'a>(
285        &'a self,
286        url: &'a str,
287        instruction: &'a str,
288    ) -> BoxFuture<'a, Result<pagesift_api::types::Observed, ApiError>> {
289        Box::pin(async move {
290            let (found, used) = Scraper::observe(self, url, instruction).await?;
291            let usage = format!("llm_usage: {} calls, {} tokens", used.calls, used.tokens);
292            Ok((found, vec![usage]))
293        })
294    }
295
296    fn find_similar<'a>(
297        &'a self,
298        url: &'a str,
299        selector: &'a str,
300        opts: &'a ScrapeOptions,
301    ) -> BoxFuture<'a, Result<pagesift_api::types::SimilarData, ApiError>> {
302        Box::pin(Scraper::find_similar(self, url, selector, opts))
303    }
304
305    fn interact<'a>(
306        &'a self,
307        id: &'a str,
308        req: &'a pagesift_api::types::InteractRequest,
309        owner: Option<&'a str>,
310    ) -> BoxFuture<'a, Result<pagesift_api::types::InteractResponse, ApiError>> {
311        Box::pin(Scraper::interact(self, id, req, owner))
312    }
313
314    fn stop_interact<'a>(
315        &'a self,
316        id: &'a str,
317        owner: Option<&'a str>,
318    ) -> BoxFuture<'a, Result<pagesift_api::types::InteractStopped, ApiError>> {
319        Box::pin(Scraper::stop_interact(self, id, owner))
320    }
321
322    fn agent_check(&self, req: &pagesift_api::types::AgentRequest) -> Result<(), ApiError> {
323        agent::check(self.llm.is_some(), self.search.is_some(), req)
324    }
325
326    fn agent<'a>(
327        &'a self,
328        req: &'a pagesift_api::types::AgentRequest,
329        sink: Arc<dyn pagesift_api::AgentSink>,
330    ) -> BoxFuture<'a, pagesift_api::types::AgentOutcome> {
331        Box::pin(async move {
332            match &self.llm {
333                Some(llm) => agent::run(self, llm, self.search.is_some(), req, sink.as_ref()).await,
334                None => pagesift_api::types::AgentOutcome {
335                    finish_reason: "error".into(),
336                    error: Some(pagesift_api::types::AgentFailure::from(
337                        &ApiError::unsupported("agent needs an LLM: set [llm] in pagesift.toml"),
338                    )),
339                    ..Default::default()
340                },
341            }
342        })
343    }
344
345    /// 统一准入：批量任务的并发取当前全局在途数目标（限速器的 USL 控制器）。
346    fn admitted_concurrency(&self) -> Option<usize> {
347        Some(self.concurrency())
348    }
349
350    fn supported_formats(&self) -> &'static [&'static str] {
351        match (self.llm.is_some(), self.store.is_some()) {
352            (true, true) => &[
353                "markdown",
354                "html",
355                "rawHtml",
356                "links",
357                "images",
358                "extract",
359                "summary",
360                "json",
361                "changeTracking",
362            ],
363            (true, false) => &[
364                "markdown", "html", "rawHtml", "links", "images", "extract", "summary", "json",
365            ],
366            (false, true) => &[
367                "markdown",
368                "html",
369                "rawHtml",
370                "links",
371                "images",
372                "extract",
373                "changeTracking",
374            ],
375            (false, false) => &["markdown", "html", "rawHtml", "links", "images", "extract"],
376        }
377    }
378}
379
380/// 放开内网时浏览器子请求的检查：只限制协议，不查地址。
381struct SchemeGuard;
382
383impl RequestGuard for SchemeGuard {
384    fn check<'a>(&'a self, req: &'a SubRequest) -> pagesift_cdp::BoxFuture<'a, Verdict> {
385        let ok = Url::parse(&req.url)
386            .is_ok_and(|u| matches!(u.scheme(), "http" | "https" | "data" | "blob" | "about"));
387        Box::pin(async move {
388            if ok {
389                Verdict::Allow
390            } else {
391                Verdict::Block("scheme not allowed".into())
392            }
393        })
394    }
395}
396
397#[cfg(test)]
398mod tests {
399    use super::*;
400
401    fn content(md: &str, published: bool, og_article: bool) -> pagesift_content::Document {
402        let mut d = pagesift_content::process("<p>x</p>", &Default::default());
403        d.markdown = md.to_string();
404        d.metadata.published_time = published.then(|| "2026-10-21T00:00:00Z".into());
405        if og_article {
406            d.metadata.extra.push(("og:type".into(), "article".into()));
407        }
408        d
409    }
410
411    #[test]
412    fn static_learning_cannot_suppress_client_code() {
413        for html in [
414            "<script>hydrate()</script>",
415            "<script src='/app.js'></script>",
416            "<SCRIPT TYPE='module' src='/view.js'></SCRIPT>",
417            "<body onload='hydrate()'>",
418            "<button onclick='hydrate()'>Open</button>",
419        ] {
420            assert!(has_client_code(html), "{html}");
421        }
422        for html in [
423            "<article><p>Short but complete.</p></article>",
424            "<script type='application/ld+json'>{}</script>",
425            "<script type='APPLICATION/JSON; charset=utf-8'>{}</script>",
426            "<script type='text/plain'>inert sample</script>",
427        ] {
428            assert!(!has_client_code(html), "{html}");
429        }
430    }
431
432    #[test]
433    fn render_rule_weights_cjk_and_trusts_short_articles() {
434        assert_eq!(text_weight("ab c\n"), 3);
435        assert_eq!(text_weight("日本語のニュース"), 8 * CJK_WEIGHT);
436        // 500 个汉字按 1500 算，不用渲染；500 个拉丁字母要渲染
437        assert!(!needs_render(&content(&"字".repeat(500), false, false)));
438        assert!(needs_render(&content(&"a".repeat(500), false, false)));
439        // 认出是文章且有 300 以上：不渲染；太短的文章照样渲染
440        assert!(!needs_render(&content(&"a".repeat(400), true, false)));
441        assert!(!needs_render(&content(&"a".repeat(400), false, true)));
442        assert!(needs_render(&content(&"a".repeat(100), true, true)));
443    }
444
445    #[test]
446    fn render_errors_are_classified() {
447        use pagesift_cdp::Error as E;
448        for (e, code) in [
449            (
450                E::Action {
451                    index: 0,
452                    message: "no element".into(),
453                },
454                ErrorCode::InvalidRequest,
455            ),
456            (E::InvalidOptions("x".into()), ErrorCode::InvalidRequest),
457            (E::ExecuteJavascriptDisabled, ErrorCode::UnsupportedOption),
458            (E::Timeout, ErrorCode::Timeout),
459            (E::Busy, ErrorCode::Busy),
460            (
461                E::Navigation("net::ERR_NAME_NOT_RESOLVED".into()),
462                ErrorCode::DnsFailure,
463            ),
464            (
465                E::Navigation("net::ERR_CONNECTION_RESET".into()),
466                ErrorCode::NetworkError,
467            ),
468            (
469                E::Navigation("net::ERR_CERT_DATE_INVALID".into()),
470                ErrorCode::TlsCertExpired,
471            ),
472            (
473                E::Navigation("net::ERR_CERT_COMMON_NAME_INVALID".into()),
474                ErrorCode::TlsHostnameMismatch,
475            ),
476            (
477                E::Navigation("net::ERR_CERT_AUTHORITY_INVALID".into()),
478                ErrorCode::TlsUntrustedRoot,
479            ),
480            (
481                E::Navigation("net::ERR_CERT_REVOKED".into()),
482                ErrorCode::TlsError,
483            ),
484            (E::ConnectionClosed, ErrorCode::UpstreamError),
485        ] {
486            assert_eq!(render_error_code(&e), code, "{e}");
487        }
488    }
489
490    #[test]
491    fn finds_only_visible_same_origin_and_srcdoc_iframes() {
492        let base = Url::parse("https://a.test/page").unwrap();
493        let html = r#"<body>
494            <iframe src="https://www.googletagmanager.com/ns.html"></iframe>
495            <iframe src="/pixel" style="display: none"></iframe>
496            <iframe src="/hidden" aria-hidden="true"></iframe>
497            <iframe src="/tiny" width="1" height="1"></iframe>
498            <iframe src="about:blank"></iframe>
499            <iframe src="/comments?id=1&amp;x=2" height="160"></iframe>
500            <iframe srcdoc="&lt;p&gt;inline&lt;/p&gt;"></iframe>
501            <iframe srcdoc="&lt;p&gt;x&lt;/p&gt;" style="visibility:hidden"></iframe>
502        </body>"#;
503        let frames = content_iframes(html, &base);
504        assert_eq!(frames.len(), 2, "{frames:?}");
505        assert_eq!(frames[0].0, "/comments?id=1&x=2");
506        assert_eq!(
507            frames[0].1.as_ref().map(Url::as_str),
508            Some("https://a.test/comments?id=1&x=2")
509        );
510        assert_eq!(frames[1], ("<p>inline</p>".to_string(), None));
511        assert!(content_iframes("<p>no frames</p>", &base).is_empty());
512    }
513
514    #[test]
515    fn replaces_the_matching_iframe_element() {
516        let mut html = r#"<p>a</p><IFRAME src="/x" hidden></IFRAME><iframe src="/x">fallback</iframe><p>b</p>"#
517            .to_string();
518        assert!(replace_iframe(&mut html, "<div>X</div>", |el| {
519            !hidden_frame(el) && src_is(el, "/x")
520        }));
521        assert_eq!(
522            html,
523            r#"<p>a</p><IFRAME src="/x" hidden></IFRAME><div>X</div><p>b</p>"#
524        );
525        assert!(!replace_iframe(&mut html, "", |el| src_is(el, "/missing")));
526    }
527
528    #[test]
529    fn takes_the_body_of_an_iframe_document() {
530        assert_eq!(
531            body_inner("<html><head><title>t</title></head><BODY class=x><p>hi</p></body></html>"),
532            "<p>hi</p>"
533        );
534        assert_eq!(body_inner("<p>fragment</p>"), "<p>fragment</p>");
535    }
536
537    #[test]
538    fn meta_keys_follow_firecrawl_style() {
539        assert_eq!(pagesift_sink::meta::key("og:image"), "ogImage");
540        assert_eq!(pagesift_sink::meta::key("og:site_name"), "ogSiteName");
541        assert_eq!(
542            pagesift_sink::meta::key("article:published_time"),
543            "articlePublishedTime"
544        );
545        assert_eq!(pagesift_sink::meta::key("dc.title"), "dcTitle");
546    }
547}
pagesift_engine/engine/mod.rs

pagesift_engine/engine/
mod.rs
