mesalogo
010aafb6f4
Main module evaluation drops from ~420 ms to ~330 ms. Types stay as import type; the DeepSeek Harness chunks stay static because the protocol chunk shares code with Main startup paths.
213 lines
11 KiB
TypeScript
213 lines
11 KiB
TypeScript
import { randomUUID } from 'node:crypto'
|
||
import { z } from 'zod'
|
||
import {
|
||
maximumDocumentExtractedCharacters,
|
||
type httpOcrSettingsSchema
|
||
} from '../shared/document-parsing-contracts'
|
||
import type { ParsedDocument, ParsedSection } from './knowledge/document-parser'
|
||
import { hasExtractedDocumentText } from './document-extracted-text'
|
||
|
||
type Settings = z.infer<typeof httpOcrSettingsSchema>
|
||
export type ParsedDocumentImage = {
|
||
id: string
|
||
pageNumber: number
|
||
key: string
|
||
locator?: string
|
||
mimeType: 'image/jpeg' | 'image/png' | 'image/webp'
|
||
width: number
|
||
height: number
|
||
data: Buffer
|
||
}
|
||
|
||
const maximumResponseBytes = 32 * 1024 * 1024
|
||
const pageSchema = z.object({
|
||
prunedResult: z.record(z.string(), z.unknown()),
|
||
markdown: z.object({
|
||
text: z.string().max(maximumDocumentExtractedCharacters),
|
||
images: z.record(z.string(), z.string()).nullish()
|
||
})
|
||
})
|
||
const responseSchema = z.object({
|
||
errorCode: z.number().optional(),
|
||
result: z.object({ layoutParsingResults: z.array(pageSchema).min(1).max(500) })
|
||
})
|
||
|
||
export class HttpDocumentOcr {
|
||
constructor(private readonly settings: Settings, private readonly apiKey?: string) {}
|
||
|
||
private async request(path: string, body?: unknown, signal?: AbortSignal): Promise<unknown> {
|
||
signal?.throwIfAborted()
|
||
if (!this.settings.baseUrl) throw new Error('请在文档解析设置中保存 HTTP OCR 服务地址')
|
||
if (this.settings.authentication === 'bearer' && !this.apiKey) {
|
||
throw new Error('HTTP OCR Bearer 凭据未配置')
|
||
}
|
||
const timeout = AbortSignal.timeout(this.settings.timeoutSeconds * 1000)
|
||
const combined = signal ? AbortSignal.any([signal, timeout]) : timeout
|
||
const response = await fetch(`${this.settings.baseUrl.replace(/\/+$/u, '')}${path}`, {
|
||
method: body === undefined ? 'GET' : 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
...(this.settings.authentication === 'bearer' ? { Authorization: `Bearer ${this.apiKey}` } : {})
|
||
},
|
||
...(body === undefined ? {} : { body: JSON.stringify(body) }),
|
||
signal: combined,
|
||
redirect: 'error'
|
||
})
|
||
if (!response.ok) {
|
||
await response.body?.cancel()
|
||
throw new Error(`HTTP OCR ${path}：HTTP ${response.status}`)
|
||
}
|
||
if (!response.body) throw new Error('HTTP OCR 返回空响应')
|
||
const reader = response.body.getReader()
|
||
const chunks: Uint8Array[] = []
|
||
let size = 0
|
||
try {
|
||
while (true) {
|
||
const chunk = await reader.read()
|
||
if (chunk.done) break
|
||
size += chunk.value.byteLength
|
||
if (size > maximumResponseBytes) throw new Error('HTTP OCR 响应超过 32 MiB 限制')
|
||
chunks.push(chunk.value)
|
||
}
|
||
} finally {
|
||
await reader.cancel()
|
||
reader.releaseLock()
|
||
}
|
||
combined.throwIfAborted()
|
||
const value: unknown = JSON.parse(Buffer.concat(chunks).toString('utf8'))
|
||
const envelope = z.object({ errorCode: z.number().optional() }).passthrough().parse(value)
|
||
if (envelope.errorCode && envelope.errorCode !== 0) {
|
||
throw new Error(`HTTP OCR ${path}：服务错误 ${envelope.errorCode}；请检查部署模型和所选高级选项`)
|
||
}
|
||
return value
|
||
}
|
||
|
||
async check(): Promise<{ checkedAt: string; declaredOptions: string[] }> {
|
||
await this.request('/health')
|
||
const schema = z.object({ components: z.object({ schemas: z.record(z.string(), z.object({ properties: z.record(z.string(), z.unknown()).optional() }).passthrough()) }).optional() })
|
||
.parse(await this.request('/openapi.json'))
|
||
return { checkedAt: new Date().toISOString(), declaredOptions: Object.keys(schema.components?.schemas.InferRequest?.properties ?? {}) }
|
||
}
|
||
|
||
async recognize(
|
||
data: Buffer,
|
||
fileType: 0 | 1,
|
||
sourcePages: number[],
|
||
signal?: AbortSignal
|
||
): Promise<Pick<ParsedDocument, 'sections' | 'images' | 'missingImages' | 'warnings' | 'restructure'>> {
|
||
const labels = this.settings.filterMode === 'default' ? undefined
|
||
: this.settings.filterMode === 'all' ? [] : this.settings.ignoredLabels
|
||
const response = responseSchema.parse(await this.request('/layout-parsing', {
|
||
file: data.toString('base64'), fileType,
|
||
...this.settings.options,
|
||
...(labels === undefined ? {} : { markdownIgnoreLabels: labels }),
|
||
returnMarkdownImages: true, visualize: false
|
||
}, signal))
|
||
const pages = response.result.layoutParsingResults
|
||
if (pages.length !== sourcePages.length) throw new Error('HTTP OCR 返回页数与输入不符，无法确定来源页码')
|
||
const warnings: string[] = []
|
||
const images: ParsedDocumentImage[] = []
|
||
const missingImages: NonNullable<ParsedDocument['missingImages']> = []
|
||
const sections: ParsedSection[] = []
|
||
let imageBytes = 0
|
||
let characters = 0
|
||
for (const [index, page] of pages.entries()) {
|
||
signal?.throwIfAborted()
|
||
const pageNumber = sourcePages[index]!
|
||
let content = page.markdown.text
|
||
const references = [...content.matchAll(/!\[[^\]]*\]\(([^)\s]+)\)|<img\b[^>]*\bsrc\s*=\s*["']([^"']+)["'][^>]*>/giu)]
|
||
for (const key of new Set(references.map((match) => match[1] ?? match[2]!))) {
|
||
signal?.throwIfAborted()
|
||
const encoded = page.markdown.images?.[key]
|
||
let replacement: string
|
||
try {
|
||
if (!encoded || /^https?:/iu.test(encoded)) throw new Error('未返回内联图片字节')
|
||
const raw = encoded.replace(/^data:image\/(?:jpeg|png|webp);base64,/iu, '')
|
||
if (!/^[A-Za-z0-9+/]+={0,2}$/u.test(raw)) throw new Error('Base64 无效')
|
||
const bytes = Buffer.from(raw, 'base64')
|
||
if (!bytes.length || bytes.length > 12 * 1024 * 1024) throw new Error('图片超过 12 MiB 或为空')
|
||
const mimeType = bytes.subarray(0, 3).equals(Buffer.from([255, 216, 255])) ? 'image/jpeg'
|
||
: bytes.subarray(0, 8).equals(Buffer.from([137, 80, 78, 71, 13, 10, 26, 10])) ? 'image/png'
|
||
: bytes.toString('ascii', 0, 4) === 'RIFF' && bytes.toString('ascii', 8, 12) === 'WEBP' ? 'image/webp' : undefined
|
||
if (!mimeType) throw new Error('图片格式不支持')
|
||
// Native canvas loads on first use (P6).
|
||
const { loadImage } = await import('@napi-rs/canvas')
|
||
const image = await loadImage(bytes)
|
||
if (!image.width || !image.height || image.width * image.height > 40_000_000) throw new Error('图片尺寸无效或超过 4000 万像素')
|
||
imageBytes += bytes.length
|
||
if (imageBytes > maximumResponseBytes) throw new Error('解码图片总量超过 32 MiB')
|
||
const id = randomUUID()
|
||
images.push({ id, pageNumber, key, mimeType, width: image.width, height: image.height, data: bytes })
|
||
replacement = ``
|
||
} catch (error) {
|
||
signal?.throwIfAborted()
|
||
missingImages.push({ pageNumber, key, reason: error instanceof Error ? error.message : '解码失败' })
|
||
const warning = `第 ${pageNumber} 页图片未保存：${error instanceof Error ? error.message : '解码失败'}`
|
||
warnings.push(warning)
|
||
replacement = `[${warning}]`
|
||
}
|
||
for (const reference of references.filter((match) => (match[1] ?? match[2]) === key)) {
|
||
content = content.replace(reference[0], replacement)
|
||
}
|
||
}
|
||
characters += content.length
|
||
if (characters > maximumDocumentExtractedCharacters) throw new Error('HTTP OCR 文字超过 5,000,000 字符')
|
||
sections.push({ locator: `第 ${pageNumber} 页`, pageNumber, sourcePages: [pageNumber], content, method: 'ocr' })
|
||
if (!content.trim()) warnings.push(`第 ${pageNumber} 页未返回文字或有效插图`)
|
||
}
|
||
// Restructuring is only meaningful for contiguous original pages.
|
||
let restructure: ParsedDocument['restructure']
|
||
if (this.settings.mergeTables || this.settings.relevelTitles) {
|
||
if (sourcePages.some((page, index) => index > 0 && page !== sourcePages[index - 1]! + 1)) {
|
||
throw new Error('不连续来源页不能进行跨页整理')
|
||
}
|
||
signal?.throwIfAborted()
|
||
try {
|
||
const result = responseSchema.parse(await this.request('/restructure-pages', {
|
||
pages: pages.map((page) => ({ prunedResult: page.prunedResult, markdownImages: page.markdown.images ?? {} })),
|
||
mergeTables: this.settings.mergeTables, relevelTitles: this.settings.relevelTitles,
|
||
...(this.settings.options.prettifyMarkdown === undefined ? {} : { prettifyMarkdown: this.settings.options.prettifyMarkdown }),
|
||
...(this.settings.options.showFormulaNumber === undefined ? {} : { showFormulaNumber: this.settings.options.showFormulaNumber }),
|
||
returnMarkdownImages: true,
|
||
concatenatePages: false
|
||
}, signal))
|
||
const changed = JSON.stringify(result.result.layoutParsingResults.map((page) => page.markdown.text)) !== JSON.stringify(pages.map((page) => page.markdown.text))
|
||
restructure = { changed, sourcePages }
|
||
if (changed) {
|
||
try {
|
||
const merged: string[] = []
|
||
for (const page of result.result.layoutParsingResults) {
|
||
let content = page.markdown.text
|
||
const references = [...content.matchAll(/!\[[^\]]*\]\(([^)\s]+)\)|<img\b[^>]*\bsrc\s*=\s*["']([^"']+)["'][^>]*>/giu)]
|
||
for (const reference of references) {
|
||
const key = reference[1] ?? reference[2]!
|
||
const encoded = page.markdown.images?.[key]
|
||
if (!encoded || /^https?:/iu.test(encoded)) throw new Error('重组图片缺少内联数据')
|
||
const bytes = Buffer.from(encoded.replace(/^data:image\/(?:jpeg|png|webp);base64,/iu, ''), 'base64')
|
||
const matches = images.filter((image) => image.data.equals(bytes))
|
||
if (matches.length !== 1) throw new Error('重组图片不能唯一对应原始页面')
|
||
content = content.replace(reference[0], `![第 ${matches[0]!.pageNumber} 页图片](asset:${matches[0]!.id})`)
|
||
}
|
||
merged.push(content)
|
||
}
|
||
const content = merged.join('\n\n')
|
||
if (!content.trim() || content.length > maximumDocumentExtractedCharacters) throw new Error('重组正文为空或超过容量限制')
|
||
// The grouped section represents all input pages, never a guessed single page.
|
||
sections.splice(0, sections.length, { locator: `跨页整理 · 来源页 ${sourcePages.join('、')}`, sourcePages, content, method: 'ocr' })
|
||
} catch (error) {
|
||
warnings.push(`跨页整理未完成，保留逐页结果：${error instanceof Error ? error.message : '来源映射失败'}`)
|
||
}
|
||
}
|
||
} catch (error) {
|
||
signal?.throwIfAborted()
|
||
warnings.push(`跨页整理请求失败，保留逐页结果：${error instanceof Error ? error.message : '请求失败'}`)
|
||
}
|
||
}
|
||
signal?.throwIfAborted()
|
||
if (!hasExtractedDocumentText(sections.map((section) => section.content).join('\n\n')) && images.length === 0) {
|
||
throw new Error('HTTP OCR 未返回文字或有效图片')
|
||
}
|
||
return { sections, images, missingImages, warnings, ...(restructure ? { restructure } : {}) }
|
||
}
|
||
}
|