tim/kangkang
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

```

Browse Source 
feat(AI): 集成MNN推理引擎替换MLX作为主AI运行时

- 引入MNN(alibaba) + Arm SME2 + CPU作为主AI运行时，支持A19/iPhone17的
  SME2和A17的NEON加速
- 添加MLX Swift作为兜底GPU推理方案，实现双后端切换机制
- 使用单一Qwen3.5-2B多模态模型(1.2GB)，替代原有的LLM+VL分离架构
- 实现InferenceEngine.current引擎选择逻辑，真机默认MNN，模拟器回退MLX
- 更新AIAgent架构，通过MNNLLMBridge(ObjC++) → MNNBackend进行推理
- 修改队列机制防止并发推理导致OOM，使用信号量闸门控制显存占用
- 更新文档中的技术栈说明、模块边界和周次交付计划
```
This commit is contained in:
link2026
2026-06-15 09:24:59 +08:00
parent 6c6a950140
commit 9d856fcfc4
37 changed files with 2605 additions and 430 deletions
Download Patch File Download Diff File Expand all files Collapse all files

325
康康/Features/Profile/MedicationScanFlow.swift
View File

@@ -2,28 +2,41 @@ import SwiftUI
import SwiftData
import UIKit
/// 「拍药盒入档」流程:拍药盒/说明书 → Vision OCR → LLM 结构化 → 核对 → 落库。
/// 入口:「+ 新建 · 健康日记 · 拍药盒」与「我的 · 个人资料 · 当前用药」。
/// 两个入口确认后都走 `MedicationArchiver`:记一条「用药」日记(进记录时间线)+ 同步当前用药。
/// 只识别入档,不做用药提醒/剂量建议(§1)。
/// 「拍药盒入库」流程:拍药盒/说明书(最多 5 张,选一张识别)→ Vision OCR → LLM 结构化 → 核对 → 存入药品库(连同原图)。
/// 入口:「记录 · 药品库」与「记录 · 健康日记 · 拍药盒」。
/// 两个入口确认后都走 `MedicationArchiver`:每条药建一个 `Medication`(挂原图),不写日记、不写当前用药。
/// 服用流水改由「写日记 · 用药」生成带 `medicationTag` 的 DiaryEntry。只识别入库,不做用药提醒/剂量建议(§1)。
///
/// 状态机(与 QuickRegionCaptureFlow 同构):
/// 状态机:
/// ```
/// idle(相机/相册) → recognizing(OCR + LLM) → confirm(核对可编辑) → onSave → 关闭
/// │ 失败/没读出 ──────► confirm(空行 + 警示文案,手动补)
/// idle(相机/相册) ─拍到第1张→ collecting(复看:删/继续拍≤5/选一张/开始识别)
/// │ 开始识别
/// ▼
/// recognizing(选中单张 OCR + LLM) ─→ confirm(核对一种药) ─onSave→ 关闭
/// │ 失败/没读出 ───────────────► confirm(空行 + 警示)
/// ```
struct MedicationScanFlow: View {
/// 用户确认后回传条目文本(非空,如 "缬沙坦胶囊 80mg · 一日一次")。落库由调用方做。
let onSave: ([String]) -> Void
/// 用户确认后回传(结构化药品, 原图)。入库由调用方做(走 MedicationArchiver.archive(medications:))。
let onSave: ([ParsedMedication], [UIImage]) -> Void
let onClose: () -> Void
/// 一种药最多关联 5 张原图(正面/背面/说明书…)。
static let maxImages = 5
@State private var phase: Phase = .idle
/// 已拍/已选的原图,跨 collecting → recognizing → confirm 一直留着,确认时全部作为该药原图落库。
@State private var images: [UIImage] = []
/// 识别用的照片索引(在多张里单选一张)。一次只记一种药 → 只 OCR 这一张;删图时校正。
@State private var recognizeIndex = 0
/// 在 collecting 复看页「继续拍/继续选」时弹相机或相册。
@State private var showMoreCapture = false
/// 识别任务句柄:识别中点「取消」要能立刻中断,不留后台推理。
@State private var recognitionTask: Task<Void, Never>?
enum Phase {
case idle
case recognizing(image: UIImage)
case collecting
case recognizing
case confirm(items: [EditableMedication], warning: String?)
}
@@ -35,6 +48,8 @@ struct MedicationScanFlow: View {
var include: Bool = true
}
private var remainingSlots: Int { max(0, Self.maxImages - images.count) }
var body: some View {
content
.background(Tj.Palette.sand.ignoresSafeArea())
@@ -45,10 +60,14 @@ struct MedicationScanFlow: View {
switch phase {
case .idle:
// 不整体 ignoresSafeArea:相机内部已全屏黑底,忽略安全区会让「取消」顶进灵动岛。
captureEntry
initialCaptureEntry
case .recognizing(let image):
recognizingView(image: image)
case .collecting:
collectingView
.fullScreenCover(isPresented: $showMoreCapture) { moreCaptureSheet }
case .recognizing:
recognizingView
case .confirm(let items, let warning):
NavigationStack {
@@ -56,7 +75,7 @@ struct MedicationScanFlow: View {
items: items,
warning: warning,
onSave: { saveItems($0) },
onRetake: { phase = .idle }
onRetake: { images = []; phase = .idle }
)
.navigationTitle("核对药品")
.navigationBarTitleDisplayMode(.inline)
@@ -72,31 +91,160 @@ struct MedicationScanFlow: View {
// MARK: - 入口:拍照(真机)/ 相册(模拟器)
/// 首张:进入即拍/选。拿到第一张就转 collecting 复看。
@ViewBuilder
private var captureEntry: some View {
private var initialCaptureEntry: some View {
#if targetEnvironment(simulator)
PhotoPickerSheet(
onFinish: { images in
if let first = images.first { startRecognition(first) } else { onClose() }
onFinish: { picked in
appendImages(picked)
if images.isEmpty { onClose() } else { phase = .collecting }
},
onCancel: onClose
)
#else
SingleShotCameraView(
onCapture: { startRecognition($0) },
onCapture: { appendImages([$0]); phase = .collecting },
onCancel: onClose
)
#endif
}
private func recognizingView(image: UIImage) -> some View {
/// collecting 复看页里「继续拍/继续选」弹出的二次采集。
@ViewBuilder
private var moreCaptureSheet: some View {
#if targetEnvironment(simulator)
PhotoPickerSheet(
onFinish: { picked in appendImages(picked); showMoreCapture = false },
onCancel: { showMoreCapture = false }
)
#else
SingleShotCameraView(
onCapture: { appendImages([$0]); showMoreCapture = false },
onCancel: { showMoreCapture = false }
)
#endif
}
private func appendImages(_ new: [UIImage]) {
guard remainingSlots > 0 else { return }
images.append(contentsOf: new.prefix(remainingSlots))
}
// MARK: - 复看(已拍 N 张:删 / 继续拍 / 开始识别)
private var collectingView: some View {
VStack(spacing: 0) {
ScrollView {
LazyVGrid(columns: [GridItem(.adaptive(minimum: 96), spacing: 12)], spacing: 12) {
ForEach(Array(images.enumerated()), id: \.offset) { idx, img in
let isPick = idx == recognizeIndex
ZStack(alignment: .topTrailing) {
Image(uiImage: img)
.resizable()
.scaledToFill()
.frame(width: 96, height: 96)
.clipShape(RoundedRectangle(cornerRadius: Tj.Radius.md, style: .continuous))
.overlay(
RoundedRectangle(cornerRadius: Tj.Radius.md, style: .continuous)
.strokeBorder(isPick ? Tj.Palette.ink : Color.clear, lineWidth: 3)
)
.overlay(alignment: .bottomLeading) {
if isPick {
Text("识别此张")
.font(.tjScaled(10, weight: .semibold))
.foregroundStyle(Tj.Palette.paper)
.padding(.horizontal, 6)
.padding(.vertical, 3)
.background(Capsule().fill(Tj.Palette.ink))
.padding(5)
}
}
// 点图把它选为「识别用」那张(单选)。
.onTapGesture { recognizeIndex = idx }
Button {
images.remove(at: idx)
// 校正识别索引:删选中前面的图要左移;删到越界则收回末尾。
if images.isEmpty {
recognizeIndex = 0
phase = .idle
} else if idx < recognizeIndex {
recognizeIndex -= 1
} else if recognizeIndex >= images.count {
recognizeIndex = images.count - 1
}
} label: {
Image(systemName: "xmark.circle.fill")
.font(.tjScaled(20))
.foregroundStyle(.white, .black.opacity(0.5))
.padding(4)
}
.buttonStyle(.plain)
}
}
if remainingSlots > 0 {
Button { showMoreCapture = true } label: {
VStack(spacing: 6) {
Image(systemName: "plus")
.font(.tjScaled(22, weight: .medium))
Text("继续拍")
.font(.tjScaled(12))
}
.foregroundStyle(Tj.Palette.text2)
.frame(width: 96, height: 96)
.background(
RoundedRectangle(cornerRadius: Tj.Radius.md, style: .continuous)
.strokeBorder(Tj.Palette.line, style: StrokeStyle(lineWidth: 1, dash: [4]))
)
}
.buttonStyle(.plain)
}
}
.padding(18)
}
VStack(spacing: 8) {
Text("已拍 \(images.count)/\(Self.maxImages) 张 · 可拍正面、背面、说明书")
.font(.tjScaled(12))
.foregroundStyle(Tj.Palette.text3)
if images.count > 1 {
Text("点照片选「识别此张」· 一次记一种药")
.font(.tjScaled(11))
.foregroundStyle(Tj.Palette.ink)
}
Text("照片与文字均不离开设备")
.font(.tjScaled(11))
.foregroundStyle(Tj.Palette.text3)
Button {
startRecognition()
} label: {
Text("开始识别")
.frame(maxWidth: .infinity)
}
.buttonStyle(TjPrimaryButton())
.disabled(images.isEmpty)
.opacity(images.isEmpty ? 0.4 : 1)
}
.padding(.horizontal, 18)
.padding(.bottom, 12)
}
.overlay(alignment: .topLeading) {
flowCancelButton { onClose() }
}
}
private var recognizingView: some View {
VStack(spacing: 18) {
Image(uiImage: image)
.resizable()
.scaledToFit()
.frame(maxHeight: 320)
.clipShape(RoundedRectangle(cornerRadius: Tj.Radius.md, style: .continuous))
.padding(.horizontal, 24)
if images.indices.contains(recognizeIndex) {
Image(uiImage: images[recognizeIndex])
.resizable()
.scaledToFit()
.frame(maxHeight: 320)
.clipShape(RoundedRectangle(cornerRadius: Tj.Radius.md, style: .continuous))
.padding(.horizontal, 24)
}
ProgressView().tint(Tj.Palette.ink)
Text("正在本地识别药品…")
.font(.tjScaled(14))
@@ -108,31 +256,37 @@ struct MedicationScanFlow: View {
.frame(maxWidth: .infinity, maxHeight: .infinity)
// 识别中也要能退出,不能让用户干等(§3.2 不卡死)
.overlay(alignment: .topLeading) {
Button {
flowCancelButton {
recognitionTask?.cancel()
onClose()
} label: {
Text("取消")
.font(.tjScaled( 16, weight: .medium))
.foregroundStyle(Tj.Palette.text)
.padding(.horizontal, 18)
.frame(minHeight: 44)
.background(Capsule().fill(Tj.Palette.paper))
.overlay(Capsule().strokeBorder(Tj.Palette.line, lineWidth: 1))
.contentShape(Capsule())
}
.buttonStyle(.plain)
.padding(.leading, 16)
.padding(.top, 8)
}
}
// MARK: - 识别(整图 OCR → LLM 结构化)
private func flowCancelButton(_ action: @escaping () -> Void) -> some View {
Button(action: action) {
Text("取消")
.font(.tjScaled(16, weight: .medium))
.foregroundStyle(Tj.Palette.text)
.padding(.horizontal, 18)
.frame(minHeight: 44)
.background(Capsule().fill(Tj.Palette.paper))
.overlay(Capsule().strokeBorder(Tj.Palette.line, lineWidth: 1))
.contentShape(Capsule())
}
.buttonStyle(.plain)
.padding(.leading, 16)
.padding(.top, 8)
}
private func startRecognition(_ image: UIImage) {
phase = .recognizing(image: image)
// MARK: - 识别(选中单张 OCR → LLM 结构化)
private func startRecognition() {
guard images.indices.contains(recognizeIndex) else { return }
phase = .recognizing
let target = images[recognizeIndex]
recognitionTask = Task {
let (items, warning) = await recognize(image)
let (items, warning) = await recognize(target)
guard !Task.isCancelled else { return } // 识别中点了取消:不再回写 phase
await MainActor.run {
// 全失败也不卡死:给一条空行让用户手填(§3.2 失败回退红线)。
@@ -148,13 +302,15 @@ struct MedicationScanFlow: View {
private func recognize(_ image: UIImage) async -> (items: [EditableMedication], warning: String?) {
do {
let text = try await OCRService.recognizeText(in: image)
let trimmed = text.trimmingCharacters(in: .whitespacesAndNewlines)
if trimmed.isEmpty {
// 一次只识别选中的这一张 → 一种药。
let text = (try? await OCRService.recognizeText(in: image))?
.trimmingCharacters(in: .whitespacesAndNewlines) ?? ""
if text.isEmpty {
return ([], String(appLoc: "没识别到文字,拍清楚一点再试"))
}
let parsed = try await MedicationScanService.shared.recognizeMedications(fromOCRText: trimmed)
let items = parsed.map {
let parsed = try await MedicationScanService.shared.recognizeMedications(fromOCRText: text)
// 一次一种药:即使识别出多条,也只取第一条。
let items = parsed.prefix(1).map {
EditableMedication(name: $0.name, strength: $0.strength, usage: $0.usage)
}
return (items, items.isEmpty ? String(appLoc: "没读出药品,可以手动填写") : nil)
@@ -172,34 +328,56 @@ struct MedicationScanFlow: View {
// MARK: - 保存
private func saveItems(_ items: [EditableMedication]) {
let entries = items
let meds = items
.filter { $0.include && !$0.name.trimmingCharacters(in: .whitespaces).isEmpty }
.map {
ParsedMedication(name: $0.name, strength: $0.strength, usage: $0.usage).entryText
ParsedMedication(name: $0.name.trimmingCharacters(in: .whitespaces),
strength: $0.strength.trimmingCharacters(in: .whitespaces),
usage: $0.usage.trimmingCharacters(in: .whitespaces))
}
onSave(entries)
// 确认后入药品库(连同原图)。空条目则不入库,由调用方据数组是否为空决定。
onSave(meds, images)
onClose()
}
}
// MARK: - 统一落库(MainActor,SwiftData 写主上下文必须由 View 侧持有的 ctx 来做,§3.1)
// MARK: - 入药品库(MainActor,SwiftData 写主上下文必须由 View 侧持有的 ctx 来做,§3.1)
/// 拍药盒确认后的统一落库,两个入口共用:
/// 1. 记一条带「用药」tag 的 DiaryEntry → 出现在「记录」时间线的「用药」分类
/// 2. 同步到 UserProfile.currentMedications(去重)→ AI 解读 / 身体档案 prompt 背景
/// 拍药盒确认后入药品库,两个入口(药品库页、写日记 · 拍药盒)共用:
/// 每条药建一个 `Medication`(挂原图),按 name+strength 软去重;**不写日记、不写 currentMedications**。
/// 服用流水改由「写日记 · 用药」生成带 `DiaryEntry.medicationTag` 的日记。
@MainActor
enum MedicationArchiver {
static func archive(entries: [String], in ctx: ModelContext) {
guard !entries.isEmpty else { return }
let diary = DiaryEntry(content: entries.joined(separator: "\n"),
tags: [DiaryEntry.medicationTag])
ctx.insert(diary)
static func archive(medications: [ParsedMedication], images: [UIImage] = [], in ctx: ModelContext) {
guard !medications.isEmpty else { return }
let profile = UserProfileStore.loadOrCreate(in: ctx)
for entry in entries where !profile.currentMedications.contains(entry) {
profile.currentMedications.append(entry)
// 原图写加密 Vault(§5/§6:落 Application Support/Vault,目录级硬件加密)。
// 多药共享同批原图时只挂「第一条新建的药」,避免同一 JPEG 被多个 Asset 引用、
// 删一条 cascade 误删另一条还在用的文件。
let savedAssets = images
.prefix(MedicationScanFlow.maxImages)
.compactMap { try? FileVault.shared.writeJPEG($0) }
let existing = (try? ctx.fetch(FetchDescriptor<Medication>())) ?? []
var attachedImages = false
for m in medications {
// 软去重:同 name+strength 已在库则只补用法 / 刷新时间,不重复建。
if let dup = existing.first(where: { $0.name == m.name && $0.strength == m.strength }) {
if dup.usage.isEmpty, !m.usage.isEmpty { dup.usage = m.usage }
dup.updatedAt = .now
continue
}
let med = Medication(name: m.name, strength: m.strength, usage: m.usage)
if !attachedImages {
for s in savedAssets {
let asset = Asset(relativePath: s.relativePath, bytes: s.bytes)
ctx.insert(asset)
med.assets.append(asset)
}
attachedImages = true
}
ctx.insert(med)
}
profile.updatedAt = .now
try? ctx.save()
}
}
@@ -231,13 +409,8 @@ private struct MedicationConfirmView: View {
ForEach($items) { $item in
Section {
HStack {
TextField(String(appLoc: "药品名,如:缬沙坦胶囊"), text: $item.name)
.foregroundStyle(Tj.Palette.text)
Toggle("", isOn: $item.include)
.labelsHidden()
.tint(Tj.Palette.ink)
}
TextField(String(appLoc: "药品名,如:缬沙坦胶囊"), text: $item.name)
.foregroundStyle(Tj.Palette.text)
TextField(String(appLoc: "规格,如:80mg×7粒"), text: $item.strength)
.foregroundStyle(Tj.Palette.text2)
TextField(String(appLoc: "用法,如:一日一次,一次一粒"), text: $item.usage)
@@ -246,12 +419,6 @@ private struct MedicationConfirmView: View {
}
Section {
Button {
items.append(.init(name: "", strength: "", usage: ""))
} label: {
Label("再加一种", systemImage: "plus.circle")
.foregroundStyle(Tj.Palette.ink)
}
Button {
onRetake()
} label: {
@@ -259,7 +426,7 @@ private struct MedicationConfirmView: View {
.foregroundStyle(Tj.Palette.ink)
}
} footer: {
Text("将记入健康日记(记录页可查),并同步到「当前用药」供 AI 解读参考。不提供任何用药建议。")
Text("一次记一种药,多张照片都会作为这种药的原图存入药品库,供查看与 AI 解读参考。不提供任何用药建议。")
}
}
.scrollContentBackground(.hidden)
@@ -267,7 +434,7 @@ private struct MedicationConfirmView: View {
Button {
onSave(items)
} label: {
Text("保存用药记录")
Text("存入药品库")
.frame(maxWidth: .infinity)
}
.buttonStyle(TjPrimaryButton())
@@ -281,5 +448,5 @@ private struct MedicationConfirmView: View {
}
#Preview {
MedicationScanFlow(onSave: { print($0) }, onClose: {})
MedicationScanFlow(onSave: { _, _ in }, onClose: {})
}