feat(Capture): 报告识别注入 Vision OCR 参考文本,提升 2B 多模态数字准确率

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

康康/AI/Prompts/VLPrompts.swift

@@ -31,12 +31,38 @@ nonisolated enum VLPrompts {
 
     /// VL 模型不知"今天"是哪天,且 few-shot 示例里写死了日期,
     /// 必须把当天日期显式注入 prompt,模型在无报告日期时才会用对正确的回退值。
-    static func reportExtraction(today: Date = .now) -> String {
+    /// ocrText 非空时把 Vision OCR 的结果作为参考文本注入 —— Vision 抄数字比
+    /// 2B 多模态读密集小字稳;版面与表格结构仍以图片为准。
+    static func reportExtraction(today: Date = .now, ocrText: String = "") -> String {
         let f = DateFormatter()
         f.locale = Locale(identifier: "en_US_POSIX")
         f.dateFormat = "yyyy-MM-dd"
         let todayStr = f.string(from: today)
-        return reportExtractionTemplate.replacingOccurrences(of: "{{TODAY}}", with: todayStr)
+        let ocrSection: String
+        if ocrText.isEmpty {
+            ocrSection = ""
+        } else {
+            ocrSection = """
+
+
+            OCR 参考文本(系统对同一报告做文字识别的结果,可能有错字、串行或漏行;版面与表格结构以图片为准,但数值、小数点以 OCR 文字更可靠):
+            \(clipOCR(ocrText))
+
+            """
+        }
+        return reportExtractionTemplate
+            .replacingOccurrences(of: "{{TODAY}}", with: todayStr)
+            .replacingOccurrences(of: "{{OCR_SECTION}}", with: ocrSection)
+    }
+
+    /// OCR 文本截断:限制进入 prompt 的体量(2B 模型上下文有限)。截到最后一个完整行。
+    static func clipOCR(_ text: String, limit: Int = 1800) -> String {
+        guard text.count > limit else { return text }
+        let clipped = String(text.prefix(limit))
+        if let lastNewline = clipped.lastIndex(of: "\n") {
+            return String(clipped[..<lastNewline]) + "\n(后续内容过长已截断)"
+        }
+        return clipped + "\n(后续内容过长已截断)"
     }
 
     private static let reportExtractionTemplate: String = #"""
@@ -84,7 +110,7 @@ JSON schema(严格):
 输入: 一份春季体检,3 项可读
 输出:
 {"title":"春季年度体检","type":"checkup","report_date":"2026-04-12","institution":"协和医院","page_count":1,"summary":"血脂偏高、其他正常","indicators":[{"name":"低密度脂蛋白","value":"3.84","unit":"mmol/L","range":"< 3.40","status":"high","source_page":1,"source_box":[0.12,0.31,0.76,0.07]},{"name":"谷丙转氨酶","value":"32","unit":"U/L","range":"9 - 50","status":"normal","source_page":1,"source_box":[0.12,0.39,0.76,0.07]},{"name":"空腹血糖","value":"5.2","unit":"mmol/L","range":"3.9 - 6.1","status":"normal","source_page":1,"source_box":[0.12,0.47,0.76,0.07]}]}
-
+{{OCR_SECTION}}
 现在请识别图片并输出 JSON:
 """#
 

康康/Services/CaptureService.swift

@@ -1,5 +1,6 @@
 import Foundation
 import UIKit
+import ImageIO
 import SwiftData
 
 /// VL 解析结果(已结构化,可直接喂 SwiftData 模型构造)。
@@ -142,11 +143,14 @@ actor CaptureService {
             throw CaptureError.modelNotReady
         }
         let urls = assets.map { FileVault.shared.rootURL.appendingPathComponent($0.relativePath) }
+        // OCR 参考(Vision 本地,<1s/页):给 2B 多模态当数字「抄写员」,降低小字误读。
+        // 任何失败都静默回退为空串,绝不阻断识别主流程(§3.2)。
+        let ocr = await Self.ocrReference(for: urls)
         let raw: String
         do {
             raw = try await AIRuntime.shared.analyzeReport(
                 imageURLs: urls,
-                prompt: VLPrompts.reportExtraction()
+                prompt: VLPrompts.reportExtraction(ocrText: ocr)
             )
         } catch {
             throw CaptureError.inferenceFailed("\(error)")
@@ -160,6 +164,20 @@ actor CaptureService {
         }
     }
 
+    /// 对 Vault 报告图逐页 OCR 拼参考文本。最多 4 页;失败/空文本返回 ""。
+    /// 用 ImageIO 直取 CGImage(不经 UIImage,避免跨 actor 传非 Sendable 引用)。
+    private static func ocrReference(for urls: [URL]) async -> String {
+        var pages: [String] = []
+        for (idx, url) in urls.prefix(4).enumerated() {
+            guard let src = CGImageSourceCreateWithURL(url as CFURL, nil),
+                  let cg = CGImageSourceCreateImageAtIndex(src, 0, nil) else { continue }
+            guard let text = try? await OCRService.recognizeText(in: cg),
+                  !text.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else { continue }
+            pages.append(urls.count > 1 ? "【第 \(idx + 1) 页】\n\(text)" : text)
+        }
+        return pages.joined(separator: "\n")
+    }
+
     // MARK: - JSON parse(static + 纯函数 → 方便单测)
 
     /// 从 VL 输出里抠出第一段合法 JSON 对象并解析。