feat: gemini to replace blip2 (#1129)

* feat: domain profile

* fix

* fix

* fix

* fix

* fix

* refactor: curl as direct engine

* fix

* wip

* fix

* fix

* fix

* fix

* fix

---------

Co-authored-by: Sha Zhou <sha.zhou@jina.ai>

backend/functions/src/cloud-functions/crawler.ts

@@ -23,6 +23,7 @@ import { FirebaseRoundTripChecker } from '../shared/services/firebase-roundtrip-
 import { JSDomControl } from '../services/jsdom';
 import { FormattedPage, md5Hasher, SnapshotFormatter } from '../services/snapshot-formatter';
 import { CurlControl } from '../services/curl';
+import { VlmControl } from '../services/vlm';
 
 export interface ExtraScrappingOptions extends ScrappingOptions {
     withIframe?: boolean | 'quoted';
@@ -57,6 +58,7 @@ export class CrawlerHost extends RPCHost {
         protected globalLogger: Logger,
         protected puppeteerControl: PuppeteerControl,
         protected curlControl: CurlControl,
+        protected vlmControl: VlmControl,
         protected jsdomControl: JSDomControl,
         protected snapshotFormatter: SnapshotFormatter,
         protected firebaseObjectStorage: FirebaseStorageBucketControl,
@@ -281,7 +283,7 @@ export class CrawlerHost extends RPCHost {
                         continue;
                     }
 
-                    const formatted = await this.snapshotFormatter.formatSnapshot(crawlerOptions.respondWith, scrapped, targetUrl, this.urlValidMs);
+                    const formatted = await this.formatSnapshot(crawlerOptions, scrapped, targetUrl, this.urlValidMs);
                     chargeAmount = this.assignChargeAmount(formatted);
                     if (crawlerOptions.tokenBudget && chargeAmount > crawlerOptions.tokenBudget) {
                         throw new BudgetExceededError(`Token budget (${crawlerOptions.tokenBudget}) exceeded, intended charge amount ${chargeAmount}.`);
@@ -311,24 +313,25 @@ export class CrawlerHost extends RPCHost {
         if (!ctx.req.accepts('text/plain') && (ctx.req.accepts('text/json') || ctx.req.accepts('application/json'))) {
             for await (const scrapped of this.cachedScrap(targetUrl, crawlOpts, crawlerOptions)) {
                 lastScrapped = scrapped;
+                if (!crawlerOptions.isEarlyReturnApplicable()) {
+                    continue;
+                }
                 if (crawlerOptions.waitForSelector || ((!scrapped?.parsed?.content || !scrapped?.title?.trim()) && !scrapped?.pdfs?.length)) {
                     continue;
                 }
 
-                const formatted = await this.snapshotFormatter.formatSnapshot(crawlerOptions.respondWith, scrapped, targetUrl, this.urlValidMs);
+                const formatted = await this.formatSnapshot(crawlerOptions, scrapped, targetUrl, this.urlValidMs);
                 chargeAmount = this.assignChargeAmount(formatted);
 
                 if (crawlerOptions.tokenBudget && chargeAmount > crawlerOptions.tokenBudget) {
                     throw new BudgetExceededError(`Token budget (${crawlerOptions.tokenBudget}) exceeded, intended charge amount ${chargeAmount}.`);
                 }
 
-                if (crawlerOptions.isEarlyReturnApplicable()) {
+                if (scrapped?.pdfs?.length && !chargeAmount) {
-                    return formatted;
+                    continue;
                 }
 
-                if (chargeAmount && scrapped?.pdfs?.length) {
+                return formatted;
-                    return formatted;
-                }
             }
 
             if (!lastScrapped) {
@@ -338,7 +341,7 @@ export class CrawlerHost extends RPCHost {
                 throw new AssertionFailureError(`No content available for URL ${targetUrl}`);
             }
 
-            const formatted = await this.snapshotFormatter.formatSnapshot(crawlerOptions.respondWith, lastScrapped, targetUrl, this.urlValidMs);
+            const formatted = await this.formatSnapshot(crawlerOptions, lastScrapped, targetUrl, this.urlValidMs);
             chargeAmount = this.assignChargeAmount(formatted);
             if (crawlerOptions.tokenBudget && chargeAmount > crawlerOptions.tokenBudget) {
                 throw new BudgetExceededError(`Token budget (${crawlerOptions.tokenBudget}) exceeded, intended charge amount ${chargeAmount}.`);
@@ -356,32 +359,35 @@ export class CrawlerHost extends RPCHost {
 
         for await (const scrapped of this.cachedScrap(targetUrl, crawlOpts, crawlerOptions)) {
             lastScrapped = scrapped;
+
+            if (!crawlerOptions.isEarlyReturnApplicable()) {
+                continue;
+            }
+
             if (crawlerOptions.waitForSelector || ((!scrapped?.parsed?.content || !scrapped?.title?.trim()) && !scrapped?.pdfs?.length)) {
                 continue;
             }
 
-            const formatted = await this.snapshotFormatter.formatSnapshot(crawlerOptions.respondWith, scrapped, targetUrl, this.urlValidMs);
+            const formatted = await this.formatSnapshot(crawlerOptions, scrapped, targetUrl, this.urlValidMs);
             chargeAmount = this.assignChargeAmount(formatted);
             if (crawlerOptions.tokenBudget && chargeAmount > crawlerOptions.tokenBudget) {
                 throw new BudgetExceededError(`Token budget (${crawlerOptions.tokenBudget}) exceeded, intended charge amount ${chargeAmount}.`);
             }
 
-            if (crawlerOptions.isEarlyReturnApplicable()) {
+            if (crawlerOptions.respondWith === 'screenshot' && Reflect.get(formatted, 'screenshotUrl')) {
-                if (crawlerOptions.respondWith === 'screenshot' && Reflect.get(formatted, 'screenshotUrl')) {
 
-                    return assignTransferProtocolMeta(`${formatted.textRepresentation}`,
+                return assignTransferProtocolMeta(`${formatted.textRepresentation}`,
-                        { code: 302, envelope: null, headers: { Location: Reflect.get(formatted, 'screenshotUrl') } }
+                    { code: 302, envelope: null, headers: { Location: Reflect.get(formatted, 'screenshotUrl') } }
-                    );
+                );
-                }
-                if (crawlerOptions.respondWith === 'pageshot' && Reflect.get(formatted, 'pageshotUrl')) {
-
-                    return assignTransferProtocolMeta(`${formatted.textRepresentation}`,
-                        { code: 302, envelope: null, headers: { Location: Reflect.get(formatted, 'pageshotUrl') } }
-                    );
-                }
-
-                return assignTransferProtocolMeta(`${formatted.textRepresentation}`, { contentType: 'text/plain', envelope: null });
             }
+            if (crawlerOptions.respondWith === 'pageshot' && Reflect.get(formatted, 'pageshotUrl')) {
+
+                return assignTransferProtocolMeta(`${formatted.textRepresentation}`,
+                    { code: 302, envelope: null, headers: { Location: Reflect.get(formatted, 'pageshotUrl') } }
+                );
+            }
+
+            return assignTransferProtocolMeta(`${formatted.textRepresentation}`, { contentType: 'text/plain', envelope: null });
         }
 
         if (!lastScrapped) {
@@ -391,7 +397,7 @@ export class CrawlerHost extends RPCHost {
             throw new AssertionFailureError(`No content available for URL ${targetUrl}`);
         }
 
-        const formatted = await this.snapshotFormatter.formatSnapshot(crawlerOptions.respondWith, lastScrapped, targetUrl, this.urlValidMs);
+        const formatted = await this.formatSnapshot(crawlerOptions, lastScrapped, targetUrl, this.urlValidMs);
         chargeAmount = this.assignChargeAmount(formatted);
         if (crawlerOptions.tokenBudget && chargeAmount > crawlerOptions.tokenBudget) {
             throw new BudgetExceededError(`Token budget (${crawlerOptions.tokenBudget}) exceeded, intended charge amount ${chargeAmount}.`);
@@ -619,6 +625,14 @@ export class CrawlerHost extends RPCHost {
             return;
         }
 
+        if (crawlOpts?.engine === ENGINE_TYPE.VLM) {
+            const finalBrowserSnapshot = await this.getFinalSnapshot(urlToCrawl, crawlOpts, crawlerOpts);
+
+            yield* this.vlmControl.fromBrowserSnapshot(finalBrowserSnapshot);
+
+            return;
+        }
+
         let cache;
 
         if (!crawlerOpts || crawlerOpts.isCacheQueryApplicable()) {
@@ -765,6 +779,10 @@ export class CrawlerHost extends RPCHost {
             crawlOpts.extraHeaders['Accept-Language'] = opts.locale;
         }
 
+        if (opts.engine?.toLowerCase() === ENGINE_TYPE.VLM) {
+            crawlOpts.favorScreenshot = true;
+        }
+
         if (opts.injectFrameScript?.length) {
             crawlOpts.injectFrameScripts = (await Promise.all(
                 opts.injectFrameScript.map((x) => {
@@ -792,6 +810,59 @@ export class CrawlerHost extends RPCHost {
         return crawlOpts;
     }
 
+    formatSnapshot(
+        crawlerOptions: CrawlerOptions,
+        snapshot: PageSnapshot & {
+            screenshotUrl?: string;
+            pageshotUrl?: string;
+        },
+        nominalUrl?: URL,
+        urlValidMs?: number
+    ) {
+        if (crawlerOptions.engine?.toLowerCase() === ENGINE_TYPE.VLM) {
+            const output: FormattedPage = {
+                title: snapshot.title,
+                content: snapshot.parsed?.textContent,
+                url: snapshot.href,
+                pageshotUrl: snapshot.pageshotUrl,
+                [Symbol.dispose]: () => undefined,
+            };
+
+            Object.defineProperty(output, 'textRepresentation', {
+                value: snapshot.parsed?.textContent,
+                enumerable: false,
+            });
+
+            return output;
+        }
+
+        return this.snapshotFormatter.formatSnapshot(crawlerOptions.respondWith, snapshot, nominalUrl, urlValidMs);
+    }
+
+    async getFinalSnapshot(url: URL, opts?: ExtraScrappingOptions, crawlerOptions?: CrawlerOptions): Promise<PageSnapshot | undefined> {
+        const it = this.cachedScrap(url, { ...opts, engine: ENGINE_TYPE.BROWSER }, crawlerOptions);
+
+        let lastSnapshot;
+        let lastError;
+        try {
+            for await (const x of it) {
+                lastSnapshot = x;
+            }
+        } catch (err) {
+            lastError = err;
+        }
+
+        if (!lastSnapshot && lastError) {
+            throw lastError;
+        }
+
+        if (!lastSnapshot) {
+            throw new AssertionFailureError(`No content available`);
+        }
+
+        return lastSnapshot;
+    }
+
     async simpleCrawl(mode: string, url: URL, opts?: ExtraScrappingOptions) {
         const it = this.cachedScrap(url, { ...opts, minIntervalMs: 500 });
 

backend/functions/src/dto/scrapping-options.ts

@@ -461,6 +461,9 @@ export class CrawlerOptions extends AutoCastable {
         if (this.injectFrameScript?.length || this.injectPageScript?.length) {
             return false;
         }
+        if (this.engine?.toLowerCase() === ENGINE_TYPE.VLM) {
+            return false;
+        }
 
         return true;
     }

backend/functions/src/services/alt-text.ts

@@ -33,9 +33,10 @@ export class AltTextService extends AsyncService {
             const resized = this.canvasService.fitImageToSquareBox(img, 1024);
             const exported = await this.canvasService.canvasToBuffer(resized, 'image/png');
 
-            const r = await this.imageInterrogator.interrogate('blip2', {
+            const r = await this.imageInterrogator.interrogate('vertex-gemini-1.5-flash-002', {
                 image: exported,
-                // prompt: `A formal caption in one sentence, concise and in the third person: HTML <img> alt text of this image. Return "**NSFW**" if you don't feel comfortable captioning it.`
+                prompt: `Yield a concise image caption sentence in third person.`,
+                system: 'You are BLIP2, an image caption model.',
             });
 
             return r.replaceAll(/[\n\"]|(\.\s*$)/g, '').trim();

backend/functions/src/services/vlm.ts

@@ -0,0 +1,64 @@
+import { AsyncService } from 'civkit/async-service';
+import { singleton } from 'tsyringe';
+
+import { PageSnapshot } from './puppeteer';
+import { Logger } from '../shared/services/logger';
+import _ from 'lodash';
+import { AssertionFailureError } from 'civkit';
+import { LLMManager } from '../shared/services/common-llm';
+
+@singleton()
+export class VlmControl extends AsyncService {
+
+    logger = this.globalLogger.child({ service: this.constructor.name });
+
+    constructor(
+        protected globalLogger: Logger,
+        protected commonLLM: LLMManager
+    ) {
+        super(...arguments);
+    }
+
+    override async init() {
+        await this.dependencyReady();
+
+        this.emit('ready');
+    }
+
+    async* fromBrowserSnapshot(snapshot?: PageSnapshot & {
+        pageshotUrl?: string,
+    }) {
+        const pageshot = snapshot?.pageshotUrl || snapshot?.pageshot;
+
+        if (!pageshot) {
+            throw new AssertionFailureError('Screenshot of the page is not available');
+        }
+
+        const it = this.commonLLM.iterRun('vertex-gemini-1.5-flash-002', {
+            prompt: [
+                typeof pageshot === 'string' ? new URL(pageshot) : pageshot,
+                `Convert this webpage screenshot into a markdown source file that does not contain HTML tags, retaining the page language and visual structures.`,
+            ],
+
+            options: {
+                system: 'You are Reader-LM-v7, an OCR model that generates Markdown source files only. No HTML, notes and chit-chats allowed',
+                stream: true
+            }
+        });
+
+        const chunks: string[] = [];
+        for await (const txt of it) {
+            chunks.push(txt);
+            const output: PageSnapshot = {
+                ...snapshot,
+                parsed: {
+                    ...snapshot?.parsed,
+                    textContent: chunks.join(''),
+                }
+            };
+            yield output;
+        }
+
+        return;
+    }
+}

thinapps-shared

@@ -1 +1 @@
-Subproject commit 439f633d464f3fd5fe288313766a43163190b60f
+Subproject commit a17e58017ee2075edeef79893fc1bf398eeb99d0