feat: add avatar payments and finance management

2026-09-08 18:44:21 +08:00
19 changed files with 57 additions and 739 deletions
@@ -1,16 +1,12 @@
 # 构建阶段：安装依赖并打包 H5
 FROM node:18-alpine AS build
 ARG APP_GIT_SHA=unknown
 ARG APP_BUILD_TIME=unknown
 WORKDIR /app
 COPY package*.json ./
 RUN npm ci
 COPY . .
 RUN printf '{"gitSha":"%s","buildTime":"%s"}\n' "$APP_GIT_SHA" "$APP_BUILD_TIME" > public/version.json
 RUN npm run build
 # 运行阶段：nginx 托管静态资源并反向代理 /api 到后端
@@ -18,11 +14,6 @@ RUN npm run build
 # 新版 nginx(>=1.31) 用 pwrite 写 pid 文件会被拦导致致命退出；1.28 用 write() 可正常启动。
 FROM nginx:1.28-alpine
 ARG APP_GIT_SHA=unknown
 ARG APP_BUILD_TIME=unknown
 LABEL org.opencontainers.image.revision=${APP_GIT_SHA} \
      org.opencontainers.image.created=${APP_BUILD_TIME}
 COPY --from=build /app/dist /usr/share/nginx/html
 # 覆盖 nginx 默认主配置（含唯一可写的 pid /tmp/nginx.pid，规避受限容器内 /run 不可写导致反复重启）
 COPY nginx.conf /etc/nginx/nginx.conf
@@ -7,13 +7,6 @@ WORKDIR /app
 COPY requirements.txt .
 RUN pip install --no-cache-dir --timeout 120 --retries 10 -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt
 ARG APP_GIT_SHA=unknown
 ARG APP_BUILD_TIME=unknown
 ENV APP_GIT_SHA=${APP_GIT_SHA} \
    APP_BUILD_TIME=${APP_BUILD_TIME}
 LABEL org.opencontainers.image.revision=${APP_GIT_SHA} \
      org.opencontainers.image.created=${APP_BUILD_TIME}
 COPY . .
 # 后端使用 SQLite（avatar.db 落在 /app 内），平铺结构以 `uvicorn main:app` 启动
@@ -1,14 +1,13 @@
 from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
 import importlib.util
 import logging
 import os
 import logging
 from apscheduler.schedulers.asyncio import AsyncIOScheduler
 from apscheduler.triggers.interval import IntervalTrigger
-from database import engine, init_db, SessionLocal
+from database import init_db, SessionLocal
 from models import Avatar, Authorization, Organization, TokenAccount, TokenPlan, User
 from fastapi.staticfiles import StaticFiles
 import routers.avatars
@@ -55,31 +54,7 @@ app.mount("/api/files", StaticFiles(directory=UPLOAD_DIR), name="knowledge-files
@app.get("/api/health")
 def health():
-    checks = _runtime_checks()
+    return ok({"status": "ok"})
    return ok({
        "status": "ok" if all(checks.values()) else "degraded",
        "gitSha": os.getenv("APP_GIT_SHA", "unknown"),
        "buildTime": os.getenv("APP_BUILD_TIME", "unknown"),
        "checks": checks,
    })
 def _runtime_checks():
    return {
        "database": _database_is_ready(),
        "uploads": os.path.isdir(UPLOAD_DIR) and os.access(UPLOAD_DIR, os.W_OK),
        "pdfOcr": importlib.util.find_spec("pymupdf") is not None,
    }
 def _database_is_ready():
    try:
        with engine.connect() as connection:
            connection.exec_driver_sql("SELECT 1")
        return True
    except Exception:
        logger.exception("Database readiness check failed")
        return False
 def seed():
@@ -5,7 +5,6 @@ pydantic
 python-multipart
 httpx
 pypdf
 PyMuPDF>=1.24,<2
 python-docx
 openpyxl
 apscheduler>=3.10
@@ -79,34 +79,6 @@ _WRITING_SYSTEM_PATTERNS = {
 _JAPANESE_KANA = re.compile(r"[\u3040-\u30ff]")
 _KOREAN_HANGUL = re.compile(r"[\uac00-\ud7af\u1100-\u11ff]")
 _LATIN_LANGUAGE_MARKERS = {
    "English": re.compile(
        r"\b(?:i|you|we|they|he|she|have|has|had|friend|who|what|where|when|why|how|"
        r"symptoms?|disease|please|can|could|would|should|is|are|was|were|the|this|that)\b",
        re.IGNORECASE,
    ),
    "French": re.compile(
        r"\b(?:je|tu|vous|nous|ils|elle|une|des|avec|pour|pourquoi|comment|bonjour|est|sont)\b",
        re.IGNORECASE,
    ),
    "Spanish": re.compile(
        r"\b(?:yo|tu|usted|nosotros|ellos|ella|una|con|para|por que|como|hola|esta|son)\b",
        re.IGNORECASE,
    ),
    "German": re.compile(
        r"\b(?:ich|du|sie|wir|eine|mit|fur|warum|wie|hallo|ist|sind|haben)\b",
        re.IGNORECASE,
    ),
    "Portuguese": re.compile(
        r"\b(?:eu|voce|nos|eles|ela|uma|com|para|porque|como|ola|esta|sao|tenho)\b",
        re.IGNORECASE,
    ),
    "Italian": re.compile(
        r"\b(?:io|tu|voi|noi|loro|una|con|per|perche|come|ciao|sono|avere)\b",
        re.IGNORECASE,
    ),
 }
 class ChatMessage(BaseModel):
    model_config = ConfigDict(populate_by_name=True)
@@ -515,67 +487,15 @@ def _qa_requires_per_turn_rendering(
    return bool(history) or _qa_requires_language_adaptation(question, answer)
-def _latin_language_name(value: str) -> str:
+def _per_turn_language_instruction() -> str:
    scores = {
        language: len(pattern.findall(value or ""))
        for language, pattern in _LATIN_LANGUAGE_MARKERS.items()
    }
    language, score = max(scores.items(), key=lambda item: item[1])
    return language if score else "the same natural language as the latest user message"
 def _turn_language_name(value: str) -> str:
    writing_system = _dominant_writing_system(value)
    return {
        "han": "Chinese",
        "japanese": "Japanese",
        "korean": "Korean",
        "cyrillic": "the same Cyrillic-script language as the latest user message",
        "arabic": "the same Arabic-script language as the latest user message",
        "hebrew": "Hebrew",
        "devanagari": "the same Devanagari-script language as the latest user message",
        "thai": "Thai",
        "greek": "Greek",
        "latin": _latin_language_name(value),
    }.get(writing_system, "the same natural language as the latest user message")
 def _per_turn_language_instruction(question: str = "") -> str:
    language = _turn_language_name(question)
    return (
-        f"MANDATORY OUTPUT LANGUAGE FOR THIS TURN: {language}. "
+        "本轮语言覆盖指令：只根据紧随其后的最新用户消息判断本轮回答语言。"
-        "Write the entire answer only in that language. This instruction overrides the languages used by "
+        "即使此前整段对话一直使用另一种语言，只要最新消息切换了语言，本轮就必须立即切换到相同语言；"
-        "conversation history, profile data, standard answers, retrieved documents, and custom prompts. "
+        "不要沿用上一轮语言。若最新消息明确指定回答语言，以该指定为准；若混用多种语言，使用其中占主导的"
-        "Translate grounded source material faithfully when necessary. Do not mention language detection, "
+        "自然语言。不要说明你检测、切换或翻译了语言。"
        "translation, or this instruction."
    )
 def _answer_requires_language_repair(question: str, answer: str) -> bool:
    question_system = _dominant_writing_system(question)
    answer_system = _dominant_writing_system(answer)
    return (
        question_system != "unknown"
        and answer_system != "unknown"
        and question_system != answer_system
    )
 def _language_repair_messages(question: str, answer: str) -> list[dict]:
    return [
        {"role": "system", "content": _per_turn_language_instruction(question)},
        {
            "role": "system",
            "content": (
                "Rewrite the supplied draft in the mandatory output language. Preserve every grounded fact, "
                "number, proper noun, uncertainty, and safety qualification. Add no new information and output "
                "only the rewritten answer."
            ),
        },
        {"role": "user", "content": answer.strip()},
    ]
 def _canonicalize_question(value: str) -> str:
    value = _normalize_question(value)
    replacements = (
@@ -808,7 +728,7 @@ def _build_prompt(
    for item in history[-MAX_HISTORY_MESSAGES:]:
        messages.append({"role": item.role, "content": item.content} if hasattr(item, "role") else item)
    # Keep the language instruction adjacent to the current turn so long histories cannot override it.
-    messages.append({"role": "system", "content": _per_turn_language_instruction(question)})
+    messages.append({"role": "system", "content": _per_turn_language_instruction()})
    messages.append({"role": "user", "content": question.strip()})
    return messages
@@ -871,41 +791,6 @@ def _call_qwen(
    return {"answer": answer.strip(), "usage": data.get("usage") or {}}
 def _call_billed_qwen(
    db: Session,
    avatar: Avatar,
    messages: list[dict],
    temperature: float,
    usage_source: str,
    model_config: ChatModelConfig,
 ) -> tuple[str, dict]:
    reservation = reserve_avatar_tokens(
        db,
        avatar,
        usage_source,
        model_config.model,
        messages,
        model_config.max_tokens,
    )
    try:
        model_result = _call_qwen(
            messages=messages,
            temperature=temperature,
            model_config=model_config,
        )
        answer = model_result["answer"]
        token_usage = settle_reservation(
            db,
            reservation,
            model_result.get("usage"),
            fallback_total=estimate_fallback_usage(messages, answer),
        )
        return answer, token_usage
    except Exception as exc:
        release_reservation(db, reservation, str(exc))
        raise
 def _iter_qwen_stream(
    messages: list[dict], temperature: float, model_config: ChatModelConfig | None = None
 ):
@@ -1014,36 +899,32 @@ def _resolve_reply(
    token_usage = None
    if model_client is not None:
        answer = model_client(messages=messages, temperature=temperature)
        if _answer_requires_language_repair(question, str(answer or "")):
            answer = model_client(
                messages=_language_repair_messages(question, str(answer)),
                temperature=0.0,
            )
    else:
        model_config = get_chat_model_config()
-        answer, token_usage = _call_billed_qwen(
+        reservation = reserve_avatar_tokens(
            db,
            avatar,
            messages,
            temperature,
            usage_source,
-            model_config,
+            model_config.model,
            messages,
            model_config.max_tokens,
        )
-        if _answer_requires_language_repair(question, answer):
+        try:
-            logger.warning(
+            model_result = _call_qwen(
-                "chat response language mismatch avatar=%s source=%s expected=%s",
+                messages=messages,
-                avatar.id,
+                temperature=temperature,
-                usage_source,
+                model_config=model_config,
                _turn_language_name(question),
            )
-            answer, token_usage = _call_billed_qwen(
+            answer = model_result["answer"]
            token_usage = settle_reservation(
                db,
-                avatar,
+                reservation,
-                _language_repair_messages(question, answer),
+                model_result.get("usage"),
-                0.0,
+                fallback_total=estimate_fallback_usage(messages, answer),
                f"{usage_source}_language_repair",
                model_config,
            )
        except Exception as exc:
            release_reservation(db, reservation, str(exc))
            raise
    answer = str(answer or "").strip()
    if image_contexts and _answer_denies_available_image(answer):
        logger.warning(
@@ -8,8 +8,7 @@ import threading
 from datetime import datetime, timezone
 from database import SessionLocal
-from models import Avatar, KnowledgeChunk, KnowledgeDoc
+from models import KnowledgeChunk, KnowledgeDoc
 from services.pdf_ocr_service import extract_scanned_pdf_text
 import embeddings
 logger = logging.getLogger(__name__)
@@ -77,23 +76,7 @@ class KnowledgeVectorizer:
            self._set_progress(db, doc, "extracting", 8)
            text = embeddings.extract_text(path, f".{doc.file_type}")
-            if doc.file_type == "pdf" and not text.strip():
+            self._set_progress(db, doc, "chunking", 22)
                avatar = db.get(Avatar, doc.avatar_id)
                if not avatar:
                    raise ValueError("文档所属分身不存在")
                def ocr_progress(done: int, total: int):
                    percent = 8 + int((done / max(1, total)) * 20)
                    self._set_progress(db, doc, "ocr", min(percent, 28))
                self._set_progress(db, doc, "ocr", 8)
                text = extract_scanned_pdf_text(
                    db,
                    avatar,
                    path,
                    on_progress=ocr_progress,
                )
            self._set_progress(db, doc, "chunking", 29)
            chunks = embeddings.chunk_text(text)
            if not chunks:
                raise ValueError("文档没有可建立索引的文字内容")
@@ -1,130 +0,0 @@
 """OCR fallback for image-only PDF knowledge documents."""
 import logging
 import os
 import time
 from typing import Callable
 from sqlalchemy.orm import Session
 from models import Avatar
 from services.chat_model_config import get_chat_model_config
 from services.token_billing import (
    estimate_fallback_usage,
    release_reservation,
    reserve_avatar_tokens,
    settle_reservation,
 )
 from services.vision_service import call_vision_model, prepare_image
 logger = logging.getLogger(__name__)
 PDF_OCR_PROMPT = (
    "请逐字转录这一页扫描文档中的全部可见文字和表格，只输出转录内容，不要解释，不要使用 Markdown 代码块。"
    "保留标题、段落、项目编号、数值和自然换行；看不清的内容写作[无法辨认]，不要猜测、纠错或补全。"
 )
 def _positive_int(name: str, default: int, minimum: int, maximum: int) -> int:
    try:
        value = int(os.getenv(name, str(default)))
    except ValueError:
        value = default
    return max(minimum, min(maximum, value))
 def extract_scanned_pdf_text(
    db: Session,
    avatar: Avatar,
    path: str,
    *,
    on_progress: Callable[[int, int], None] | None = None,
 ) -> str:
    """Render and OCR an image-only PDF while preserving page order."""
    try:
        import pymupdf
    except ImportError as exc:
        raise RuntimeError("扫描型 PDF 识别组件未安装") from exc
    max_pages = _positive_int("KNOWLEDGE_PDF_OCR_MAX_PAGES", 80, 1, 300)
    render_dpi = _positive_int("KNOWLEDGE_PDF_OCR_DPI", 144, 96, 200)
    max_attempts = _positive_int("KNOWLEDGE_PDF_OCR_ATTEMPTS", 3, 1, 5)
    model_config = get_chat_model_config()
    model = model_config.ocr_model or model_config.vision_model
    if not model_config.api_key or not model:
        raise RuntimeError("扫描型 PDF 需要配置视觉 OCR 模型")
    texts: list[str] = []
    with pymupdf.open(path) as document:
        total_pages = document.page_count
        if total_pages <= 0:
            raise ValueError("PDF 没有可识别页面")
        if total_pages > max_pages:
            raise ValueError(
                f"扫描型 PDF 共 {total_pages} 页，超过单次 OCR 上限 {max_pages} 页，请拆分后上传"
            )
        scale = render_dpi / 72
        for page_index in range(total_pages):
            page = document.load_page(page_index)
            pixmap = page.get_pixmap(
                matrix=pymupdf.Matrix(scale, scale),
                colorspace=pymupdf.csRGB,
                alpha=False,
            )
            prepared = prepare_image(pixmap.tobytes("jpeg", jpg_quality=88))
            estimate_messages = [{
                "role": "user",
                "content": f"[扫描 PDF 第 {page_index + 1}/{total_pages} 页]\n{PDF_OCR_PROMPT}",
            }]
            reservation = reserve_avatar_tokens(
                db,
                avatar,
                "knowledge_pdf_ocr",
                model,
                estimate_messages,
                model_config.vision_max_tokens,
            )
            try:
                result = None
                for attempt in range(1, max_attempts + 1):
                    try:
                        result = call_vision_model(
                            prepared,
                            model_config,
                            model=model,
                            prompt=PDF_OCR_PROMPT,
                            json_output=False,
                        )
                        break
                    except RuntimeError:
                        if attempt == max_attempts:
                            raise
                        time.sleep(min(4, attempt))
                content = str((result or {}).get("content") or "").strip()
                if not content:
                    raise RuntimeError("扫描型 PDF 页面识别结果为空")
                settle_reservation(
                    db,
                    reservation,
                    (result or {}).get("usage"),
                    fallback_total=estimate_fallback_usage(estimate_messages, content),
                )
            except Exception as exc:
                release_reservation(db, reservation, str(exc))
                raise RuntimeError(
                    f"扫描型 PDF 第 {page_index + 1}/{total_pages} 页识别失败：{exc}"
                ) from exc
            texts.append(f"[第 {page_index + 1} 页]\n{content}")
            if on_progress:
                on_progress(page_index + 1, total_pages)
            logger.info(
                "Scanned PDF OCR completed avatar=%s page=%s/%s",
                avatar.id,
                page_index + 1,
                total_pages,
            )
    return "\n\n".join(texts).strip()
@@ -6,7 +6,6 @@ from fastapi import HTTPException
 from models import Avatar, User
 from routers.chat import (
    _answer_requires_language_repair,
    _build_prompt,
    _iter_text_chunks,
    _match_standard_qa,
@@ -15,7 +14,6 @@ from routers.chat import (
    _qa_requires_per_turn_rendering,
    _require_owned_avatar,
    _resolve_reply,
    _turn_language_name,
 )
@@ -109,8 +107,8 @@ class ChatOrchestrationTests(unittest.TestCase):
        messages = fake_model.call_args.kwargs["messages"]
        self.assertEqual(messages[-1], {"role": "user", "content": "Quelle est votre adresse ?"})
        self.assertEqual(messages[-2]["role"], "system")
-        self.assertIn("MANDATORY OUTPUT LANGUAGE", messages[-2]["content"])
+        self.assertIn("本轮语言覆盖指令", messages[-2]["content"])
-        self.assertIn("French", messages[-2]["content"])
+        self.assertIn("不要沿用上一轮语言", messages[-2]["content"])
    def test_latest_user_message_has_an_adjacent_language_override(self):
        history = [
@@ -121,37 +119,8 @@ class ChatOrchestrationTests(unittest.TestCase):
        self.assertEqual(messages[-1], {"role": "user", "content": "What can you help me with?"})
        self.assertEqual(messages[-2]["role"], "system")
-        self.assertIn("MANDATORY OUTPUT LANGUAGE", messages[-2]["content"])
+        self.assertIn("最新用户消息", messages[-2]["content"])
-        self.assertIn("English", messages[-2]["content"])
+        self.assertIn("立即切换到相同语言", messages[-2]["content"])
    def test_reported_alzheimer_question_is_explicitly_english(self):
        question = "I have a friend who has symptoms of Alzheimer's disease"
        self.assertEqual(_turn_language_name(question), "English")
        messages = _build_prompt(self.avatar, [], question, [])
        self.assertIn("MANDATORY OUTPUT LANGUAGE FOR THIS TURN: English", messages[-2]["content"])
    def test_non_stream_reply_repairs_a_wrong_writing_system_before_sending(self):
        question = "I have a friend who has symptoms of Alzheimer's disease"
        fake_model = Mock(side_effect=["建议尽快就医评估。", "Please arrange a medical assessment soon."])
        result = _resolve_reply(
            None,
            self.avatar,
            question,
            [],
            qa_pairs=[],
            search_fn=lambda *_args, **_kwargs: [],
            model_client=fake_model,
            usage_source="takeover",
        )
        self.assertEqual(result["answer"], "Please arrange a medical assessment soon.")
        self.assertEqual(fake_model.call_count, 2)
        repair_messages = fake_model.call_args.kwargs["messages"]
        self.assertIn("English", repair_messages[0]["content"])
        self.assertIn("建议尽快就医评估", repair_messages[-1]["content"])
        self.assertTrue(_answer_requires_language_repair(question, "建议尽快就医评估。"))
    def test_conversational_paraphrase_matches_standard_qa(self):
        for question in ("请问一下，你们公司在哪里呀？", "请问去你们那边怎么走"):
@@ -1,33 +0,0 @@
 import main
 def test_health_reports_release_and_runtime_capabilities(monkeypatch):
    monkeypatch.setenv("APP_GIT_SHA", "test-sha")
    monkeypatch.setenv("APP_BUILD_TIME", "2026-09-09T00:00:00Z")
    monkeypatch.setattr(main, "_runtime_checks", lambda: {
        "database": True,
        "uploads": True,
        "pdfOcr": True,
    })
    response = main.health()
    assert response["code"] == 200
    assert response["data"]["status"] == "ok"
    assert response["data"]["gitSha"] == "test-sha"
    assert response["data"]["buildTime"] == "2026-09-09T00:00:00Z"
    assert response["data"]["checks"] == {
        "database": True,
        "uploads": True,
        "pdfOcr": True,
    }
 def test_health_is_degraded_when_a_required_capability_is_missing(monkeypatch):
    monkeypatch.setattr(main, "_runtime_checks", lambda: {
        "database": True,
        "uploads": True,
        "pdfOcr": False,
    })
    assert main.health()["data"]["status"] == "degraded"
@@ -238,53 +238,6 @@ def test_background_vectorizer_keeps_failure_reason_for_retry(
        db.close()
 def test_background_vectorizer_uses_ocr_for_image_only_pdf(
    tmp_path: Path,
    authorization_context,
 ):
    context = authorization_context
    with (
        patch("routers.knowledge.UPLOAD_DIR", str(tmp_path)),
        patch("routers.knowledge.knowledge_vectorizer.enqueue"),
    ):
        response = client.post(
            f"/api/avatar/{context['avatar'].id}/knowledge/docs",
            headers=context["owner_headers"],
            files={"file": ("scanned.pdf", b"image-only-pdf", "application/pdf")},
        )
    payload = response.json()["data"]
    progress = []
    with (
        patch("services.knowledge_vectorizer.UPLOAD_DIR", str(tmp_path)),
        patch("services.knowledge_vectorizer.embeddings.extract_text", return_value=""),
        patch(
            "services.knowledge_vectorizer.extract_scanned_pdf_text",
            side_effect=lambda _db, _avatar, _path, on_progress: (
                on_progress(1, 2), on_progress(2, 2), "扫描页文字"
            )[-1],
        ) as ocr,
        patch("services.knowledge_vectorizer.embeddings.embed", return_value=[[1.0, 0.0]]),
        patch.object(knowledge_vectorizer, "_set_progress", wraps=knowledge_vectorizer._set_progress) as set_progress,
    ):
        knowledge_vectorizer.vectorize_document(payload["id"])
        progress = [(call.args[2], call.args[3]) for call in set_progress.call_args_list]
    db = SessionLocal()
    try:
        stored = db.query(KnowledgeDoc).filter(KnowledgeDoc.id == payload["id"]).one()
        assert stored.status == "ready"
        assert stored.chunk_count == 1
        assert ("ocr", 18) in progress
        assert ("ocr", 28) in progress
        ocr.assert_called_once()
        db.query(KnowledgeChunk).filter(KnowledgeChunk.doc_id == stored.id).delete()
        db.delete(stored)
        db.commit()
    finally:
        db.close()
 def test_retry_queues_a_failed_document_again(
    tmp_path: Path,
    authorization_context,
@@ -1,104 +0,0 @@
 from types import SimpleNamespace
 from unittest.mock import MagicMock, patch
 import pytest
 from services.pdf_ocr_service import extract_scanned_pdf_text
 class FakePixmap:
    def tobytes(self, *_args, **_kwargs):
        return b"jpeg-page"
 class FakePage:
    def get_pixmap(self, **_kwargs):
        return FakePixmap()
 class FakeDocument:
    page_count = 2
    def __enter__(self):
        return self
    def __exit__(self, *_args):
        return None
    def load_page(self, _index):
        return FakePage()
 def test_scanned_pdf_ocr_preserves_page_order_and_reports_progress(monkeypatch):
    fake_pymupdf = SimpleNamespace(
        open=lambda _path: FakeDocument(),
        Matrix=lambda x, y: (x, y),
        csRGB="rgb",
    )
    monkeypatch.setitem(__import__("sys").modules, "pymupdf", fake_pymupdf)
    progress = []
    reservation = SimpleNamespace()
    config = SimpleNamespace(
        api_key="configured",
        ocr_model="qwen-vl-ocr",
        vision_model="vision",
        vision_max_tokens=2048,
    )
    with (
        patch("services.pdf_ocr_service.get_chat_model_config", return_value=config),
        patch("services.pdf_ocr_service.prepare_image", return_value=SimpleNamespace()),
        patch(
            "services.pdf_ocr_service.call_vision_model",
            side_effect=[
                {"content": "第一页文字", "usage": {"total_tokens": 10}},
                {"content": "第二页文字", "usage": {"total_tokens": 12}},
            ],
        ),
        patch("services.pdf_ocr_service.reserve_avatar_tokens", return_value=reservation) as reserve,
        patch("services.pdf_ocr_service.settle_reservation") as settle,
    ):
        text = extract_scanned_pdf_text(
            MagicMock(),
            SimpleNamespace(id="avatar-1"),
            "/tmp/scanned.pdf",
            on_progress=lambda done, total: progress.append((done, total)),
        )
    assert text == "[第 1 页]\n第一页文字\n\n[第 2 页]\n第二页文字"
    assert progress == [(1, 2), (2, 2)]
    assert reserve.call_count == 2
    assert settle.call_count == 2
 def test_scanned_pdf_ocr_releases_tokens_after_retries_fail(monkeypatch):
    fake_document = FakeDocument()
    fake_document.page_count = 1
    fake_pymupdf = SimpleNamespace(
        open=lambda _path: fake_document,
        Matrix=lambda x, y: (x, y),
        csRGB="rgb",
    )
    monkeypatch.setitem(__import__("sys").modules, "pymupdf", fake_pymupdf)
    monkeypatch.setenv("KNOWLEDGE_PDF_OCR_ATTEMPTS", "2")
    reservation = SimpleNamespace()
    config = SimpleNamespace(
        api_key="configured",
        ocr_model="qwen-vl-ocr",
        vision_model="vision",
        vision_max_tokens=2048,
    )
    with (
        patch("services.pdf_ocr_service.get_chat_model_config", return_value=config),
        patch("services.pdf_ocr_service.prepare_image", return_value=SimpleNamespace()),
        patch("services.pdf_ocr_service.call_vision_model", side_effect=RuntimeError("timeout")) as call,
        patch("services.pdf_ocr_service.reserve_avatar_tokens", return_value=reservation),
        patch("services.pdf_ocr_service.release_reservation") as release,
        patch("services.pdf_ocr_service.time.sleep"),
    ):
        with pytest.raises(RuntimeError, match="第 1/1 页识别失败"):
            extract_scanned_pdf_text(MagicMock(), SimpleNamespace(id="avatar-1"), "/tmp/scanned.pdf")
    assert call.call_count == 2
    release.assert_called_once()
@@ -1,56 +1,42 @@
 # 会会数字分身 —— Docker 测试实例（独立端口，不干扰现有 :8088 huihui 部署）
 services:
  avatar-backend:
-    build:
+    build: ./backend
-      context: ./backend
+    image: avatar-test-backend:latest
      args:
        APP_GIT_SHA: ${APP_GIT_SHA:?APP_GIT_SHA must be the full release commit}
        APP_BUILD_TIME: ${APP_BUILD_TIME:?APP_BUILD_TIME must be set}
    image: avatar-test-backend:${APP_GIT_SHA}
    container_name: avatar-test-backend
    restart: unless-stopped
    env_file:
      - .env
    environment:
-      DATABASE_URL: sqlite:////data/db/avatar.db
+      DATABASE_URL: sqlite:////data/avatar.db
      UPLOAD_DIR: /data/uploads
      CHAT_MODEL_CONFIG_URL: http://host.docker.internal:8000/api/ai-models/runtime/digital-avatar
    extra_hosts:
      - "host.docker.internal:host-gateway"
    volumes:
-      # Mount the directory, not only avatar.db: SQLite WAL/SHM files must survive recreation.
+      - avatar-data:/data
      - ${AVATAR_DB_DIR:?AVATAR_DB_DIR must contain the persistent avatar.db}:/data/db
      - ${AVATAR_UPLOAD_DIR:?AVATAR_UPLOAD_DIR must point to persistent uploads}:/data/uploads
    expose:
      - "8000"
    ports:
      - "8011:8000"   # 仅用于直接调试 API；前端经内部网络访问，不走 host 端口
    healthcheck:
      test: ["CMD", "python", "-c", "import json,urllib.request; d=json.load(urllib.request.urlopen('http://127.0.0.1:8000/api/health', timeout=5))['data']; assert d['status']=='ok' and all(d['checks'].values())"]
      interval: 10s
      timeout: 8s
      retries: 12
      start_period: 20s
    networks:
      - avatar-net
  avatar-frontend:
-    build:
+    build: .
-      context: .
+    image: avatar-test-frontend:latest
      args:
        APP_GIT_SHA: ${APP_GIT_SHA:?APP_GIT_SHA must be the full release commit}
        APP_BUILD_TIME: ${APP_BUILD_TIME:?APP_BUILD_TIME must be set}
    image: avatar-test-frontend:${APP_GIT_SHA}
    container_name: avatar-test-frontend
    restart: unless-stopped
    ports:
      - "8099:80"     # 浏览器访问 http://<host>:8099
    depends_on:
-      avatar-backend:
+      - avatar-backend
        condition: service_healthy
    networks:
      - avatar-net
 networks:
  avatar-net:
    driver: bridge
 volumes:
  avatar-data:
@@ -61,9 +61,7 @@ WECHAT_VIRTUAL_PRODUCT_2=<100元套餐商品ID>
 WECHAT_VIRTUAL_PRODUCT_3=<1000元套餐商品ID>
 WECHAT_VIRTUAL_PRODUCT_4=<10000元套餐商品ID>
-AVATAR_DB_DIR=/srv/digital-avatar/data/db
+DATABASE_URL=sqlite:////data/avatar.db
 AVATAR_UPLOAD_DIR=/srv/digital-avatar/data/uploads
 DATABASE_URL=sqlite:////data/db/avatar.db
 UPLOAD_DIR=/data/uploads
 CHAT_MODEL_CONFIG_URL=http://<huihuisquare-api>/api/ai-models/runtime/digital-avatar
 EMBEDDING_API_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
@@ -76,17 +74,13 @@ VISION_OCR_MODEL=qwen-vl-ocr
 VISION_MAX_OUTPUT_TOKENS=2048
 VISION_TIMEOUT_SECONDS=90
 VISION_TOKEN_RESERVE=12000
 APP_GIT_SHA=<本次发布的完整提交SHA>
 APP_BUILD_TIME=<UTC ISO-8601构建时间>
 CHAT_IMAGE_MAX_BYTES=8388608
 CHAT_IMAGE_MAX_PIXELS=16000000
 CHAT_ATTACHMENT_RETENTION_HOURS=24
 CHAT_ATTACHMENT_CLEANUP_MINUTES=60
 ```
-如生产 AI 配置中心不可用，还应提供当前项目支持的 `OPENAI_API_KEY`、`OPENAI_BASE_URL`、`CHAT_MODEL` 等兜底配置。数据库文件与上传目录必须从宿主机显式挂载，不能存放在容器临时层。
+如生产 AI 配置中心不可用，还应提供当前项目支持的 `OPENAI_API_KEY`、`OPENAI_BASE_URL`、`CHAT_MODEL` 等兜底配置。`/data` 必须挂载持久卷，数据库与知识库文件不可存放在容器临时层。
 `AVATAR_DB_DIR` 和 `AVATAR_UPLOAD_DIR` 必须是已备份的宿主机绝对路径，编排缺少任一变量都会直接拒绝构建或启动，防止误挂空卷造成用户、分身或知识库“丢失”的假象。SQLite 必须挂载整个数据库目录，不能只挂载 `avatar.db` 单文件，否则 `avatar.db-wal` 和 `avatar.db-shm` 会留在容器临时层，换容器后可能出现数据状态回退。
 `EMBEDDING_API_URL` 同时支持 OpenAI 兼容基础地址（如上面的 `/v1`）和完整的 `/v1/embeddings` 地址，后端会统一请求 `/embeddings`。发布后必须在后端容器内执行一次最小向量探针，确认返回向量数量和维度，而不能只检查 `/api/health`。
@@ -103,7 +97,7 @@ App 与 H5 积分充值使用会会支付体系的 `payment-v3/payment/pay`，
 ```bash
 BACKUP_DIR="backups/$(date +%Y%m%d-%H%M%S)"
 mkdir -p "$BACKUP_DIR"
-cp /srv/digital-avatar/data/db/avatar.db "$BACKUP_DIR/"
+cp /srv/digital-avatar/data/avatar.db "$BACKUP_DIR/"
 tar -C /srv/digital-avatar/data -czf "$BACKUP_DIR/uploads.tgz" uploads
 ```
@@ -113,22 +107,13 @@ tar -C /srv/digital-avatar/data -czf "$BACKUP_DIR/uploads.tgz" uploads
 git fetch origin
 git checkout <已验收的提交SHA>
 cd digital-avatar-app
-export APP_GIT_SHA="$(git rev-parse HEAD)"
+docker compose build --pull avatar-backend avatar-frontend
-export APP_BUILD_TIME="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
+docker compose up -d avatar-backend avatar-frontend
 docker compose build --pull --no-cache avatar-backend avatar-frontend
 docker compose up -d --force-recreate --wait avatar-backend avatar-frontend
 docker compose ps
-python3 scripts/verify-deployment.py \
+curl -fsS http://127.0.0.1:8099/api/health
  https://digital.99hui.com "$APP_GIT_SHA" \
  --backend-container avatar-backend \
  --frontend-container avatar-frontend \
  --expected-db-source /srv/digital-avatar/data/db \
  --expected-upload-source /srv/digital-avatar/data/uploads
 docker compose exec avatar-backend python -c 'import embeddings; v=embeddings.embed(["部署向量探针"]); print(len(v), len(v[0]))'
 ```
 Jenkins 必须以 `verify-deployment.py` 返回成功作为发布成功条件，不能只以镜像构建或容器启动成功作为条件。脚本会同时核对公网前后端 Git SHA、数据库可读、上传目录可写、PDF OCR 依赖和宿主机数据挂载；任意一项不一致都会返回非零状态并阻止发布标绿。镜像使用 Git SHA 标签，不再依赖可被旧缓存覆盖的 `latest`。
 生产编排应把示例中的测试端口改为内网暴露，由统一 HTTPS 网关接入。后端暂时使用 SQLite，必须保持单实例写入；若扩展为多后端实例，应先迁移到 PostgreSQL，并把延迟接管任务改为共享队列。
 ## 4. 网关要求
@@ -168,7 +153,7 @@ location /api/ {
 5. 使用过期或伪造 token 时进入登录页并显示凭证失效，不得继续访问旧用户数据。
 6. 分身聊天 SSE 逐段输出正常，Markdown 正常渲染，知识库优先级和积分扣费正常。
 7. 开启 BOXIM 主动接管后保持在线，默认三分钟回复、自定义等待时间、已读回执、分身防回环和主人发言暂停均正常。
-8. 重建容器后数据库、头像、知识库文档仍存在，`/api/health` 的 `gitSha` 与发布 SHA 一致，`database`、`uploads`、`pdfOcr` 三项检查均为 `true`。
+8. 重建容器后数据库、头像、知识库文档仍存在，`/api/health` 返回成功。
 9. `https://digital.99hui.com/api/health` 可访问，证书域名和有效期正确，HTTP 自动跳转 HTTPS。
 10. 微信和支付宝各创建一笔最小套餐订单，未付款时积分不变；支付成功后回调到账一次，重复回调积分不重复增加。
 11. 微信虚拟支付在沙箱环境完成下单、支付回调、查单兜底和退款回调；错误 OpenID、商品、环境或金额均被拒绝。
@@ -1,107 +0,0 @@
 #!/usr/bin/env python3
 """Fail a deployment unless frontend and backend run the expected release."""
 import argparse
 import json
 import os
 import subprocess
 import sys
 import urllib.request
 def fetch_json(url):
    with urllib.request.urlopen(url, timeout=20) as response:
        if response.status != 200:
            raise RuntimeError(f"{url} returned HTTP {response.status}")
        return json.load(response)
 def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("base_url", help="Public site URL, for example https://digital.99hui.com")
    parser.add_argument("expected_sha", help="Full Git commit SHA being deployed")
    parser.add_argument("--backend-container", help="Backend container name for image and mount checks")
    parser.add_argument("--frontend-container", help="Frontend container name for image checks")
    parser.add_argument("--expected-db-source", help="Required host directory mounted for SQLite and its WAL files")
    parser.add_argument("--expected-upload-source", help="Required host source mounted as the upload directory")
    args = parser.parse_args()
    base_url = args.base_url.rstrip("/")
    errors = []
    try:
        health = fetch_json(f"{base_url}/api/health").get("data") or {}
    except Exception as exc:
        errors.append(f"cannot read backend release metadata: {exc}")
        health = {}
    try:
        frontend = fetch_json(f"{base_url}/version.json")
    except Exception as exc:
        errors.append(f"cannot read frontend release metadata: {exc}")
        frontend = {}
    if health.get("status") != "ok":
        errors.append(f"backend status is {health.get('status')!r}")
    failed_checks = [name for name, passed in (health.get("checks") or {}).items() if not passed]
    if failed_checks:
        errors.append("backend checks failed: " + ", ".join(failed_checks))
    if health.get("gitSha") != args.expected_sha:
        errors.append(f"backend SHA is {health.get('gitSha')!r}")
    if frontend.get("gitSha") != args.expected_sha:
        errors.append(f"frontend SHA is {frontend.get('gitSha')!r}")
    if args.backend_container:
        backend = inspect_container(args.backend_container, errors)
        check_container_revision(backend, args.expected_sha, "backend", errors)
        check_mount(backend, args.expected_db_source, "database", errors)
        check_mount(backend, args.expected_upload_source, "uploads", errors)
    elif args.expected_db_source or args.expected_upload_source:
        errors.append("--backend-container is required when checking data mounts")
    if args.frontend_container:
        frontend_container = inspect_container(args.frontend_container, errors)
        check_container_revision(frontend_container, args.expected_sha, "frontend", errors)
    if errors:
        print("Deployment verification failed:", file=sys.stderr)
        for error in errors:
            print(f"- {error}", file=sys.stderr)
        return 1
    print(f"Deployment verified: {args.expected_sha}")
    print("Backend checks: database, uploads, pdfOcr")
    return 0
 def inspect_container(name, errors):
    try:
        output = subprocess.check_output(
            ["docker", "inspect", name], universal_newlines=True
        )
        return json.loads(output)[0]
    except Exception as exc:
        errors.append(f"cannot inspect container {name!r}: {exc}")
        return {}
 def check_container_revision(container, expected_sha, label, errors):
    actual = ((container.get("Config") or {}).get("Labels") or {}).get(
        "org.opencontainers.image.revision"
    )
    if actual != expected_sha:
        errors.append(f"{label} container image SHA is {actual!r}")
 def check_mount(container, expected_source, label, errors):
    if not expected_source:
        return
    expected = os.path.realpath(expected_source)
    sources = {
        os.path.realpath(mount.get("Source", ""))
        for mount in container.get("Mounts") or []
    }
    if expected not in sources:
        errors.append(f"{label} mount source {expected!r} is not attached")
 if __name__ == "__main__":
    raise SystemExit(main())
@@ -1,7 +1,6 @@
 import { createRouter, createWebHashHistory } from 'vue-router'
 import type { RouteRecordRaw } from 'vue-router'
 import { getAuthToken } from '@/api'
 import { isInUniWebView } from '@/utils/uniapp-bridge'
 const routes: RouteRecordRaw[] = [
  {
@@ -56,7 +55,7 @@ const routes: RouteRecordRaw[] = [
    path: '/token/charge',
    name: 'TokenCharge',
    component: () => import('@/views/TokenCharge.vue'),
-    meta: { title: '积分充值', requiresAuth: true, requiresUniWebView: true }
+    meta: { title: '积分充值', requiresAuth: true }
  },
  {
    path: '/avatar/card',
@@ -134,10 +133,6 @@ const router = createRouter({
 router.beforeEach((to, from, next) => {
  document.title = to.meta.title as string || '会会数字分身'
  if (to.meta.requiresUniWebView && !isInUniWebView()) {
    next({ path: '/avatar/manage' })
    return
  }
  const hasLocalSession = Boolean(localStorage.getItem('hh_app_token'))
  const hasInjectedSession = Boolean(getAuthToken())
  if (to.meta.requiresAuth && !hasLocalSession && !hasInjectedSession) {
@@ -12,10 +12,9 @@ export interface UniLaunchParams {
  nickname?: string
  avatar?: string
  ts?: string
  nativeShell?: string
 }
-const PARAM_KEYS: (keyof UniLaunchParams)[] = ['token', 'userId', 'nickname', 'avatar', 'ts', 'nativeShell']
+const PARAM_KEYS: (keyof UniLaunchParams)[] = ['token', 'userId', 'nickname', 'avatar', 'ts']
 function readParams(search: string, target: UniLaunchParams): void {
  const sp = new URLSearchParams(search)
@@ -25,16 +24,6 @@ function readParams(search: string, target: UniLaunchParams): void {
  }
 }
 function hasNativeShellMarker(): boolean {
  const params = getLaunchParams()
  if (params.nativeShell === 'uniapp') return true
  // Compatibility for already-installed shells. They have always appended a
  // timestamp together with the native SSO context, even before the explicit
  // nativeShell marker existed.
  return Boolean(params.ts && (params.token || params.userId))
 }
 // 是否运行在 uniapp web-view 环境中
 export function isInUniWebView(): boolean {
  const runtime = window as any
@@ -51,10 +40,7 @@ export function isInUniWebView(): boolean {
    runtime.swan?.webView ||
    runtime.tt?.miniProgram
  )
-  // `plus` can be injected after the H5 entry point runs. The native shell
+  return Boolean(runtime.uni?.webView && (isDCloudApp || isMiniProgram))
  // therefore adds a URL marker while creating its web-view URL, so the
  // payment entry does not disappear during that startup window.
  return Boolean(runtime.uni?.webView && (isDCloudApp || isMiniProgram || hasNativeShellMarker()))
 }
 // 解析 web-view 加载 URL 时原生注入的参数（token / 会会用户）
@@ -20,7 +20,7 @@
      </div>
    </section>
-    <!-- 积分余额条：仅在 uni-app 原生壳内开放充值购买。 -->
+    <!-- 积分余额条：暂时隐藏，保留完整实现便于后续恢复。 -->
    <section v-if="SHOW_POINTS_BALANCE_CARD" class="token-section">
      <div class="token-card">
        <div class="token-info">
@@ -28,7 +28,7 @@
          <span class="token-amount">{{ tokenBalance.toLocaleString() }}</span>
          <span class="token-used">累计使用 {{ tokenConsumed.toLocaleString() }}</span>
        </div>
-        <button class="recharge-btn" @click="goToRecharge">充值购买</button>
+        <button class="recharge-btn" @click="goToRecharge">充值</button>
      </div>
    </section>
@@ -89,15 +89,14 @@ import { useAvatarStore } from '@/store/avatar'
 import { useUserStore } from '@/store/user'
 import { createAvatarShareLink } from '@/api'
 import { isHuihuiEmbeddedMode } from '@/utils/embed-mode'
 import { isInUniWebView } from '@/utils/uniapp-bridge'
 const router = useRouter()
 const avatarStore = useAvatarStore()
 const userStore = useUserStore()
 const isEmbedded = isHuihuiEmbeddedMode()
-// 充值购买只在 uni-app 原生壳内提供，避免普通 H5 进入支付链路。
+// 临时产品开关：余额卡片代码保留，后续改为 true 即可恢复展示。
-const SHOW_POINTS_BALANCE_CARD = isInUniWebView()
+const SHOW_POINTS_BALANCE_CARD = false
 // 当前登录会会用户的资料（头像 / 昵称）
 const me = computed(() => userStore.user)
@@ -147,7 +147,7 @@ const documentState = (doc: any) => {
  if (['uploaded', 'parsing'].includes(String(doc.status || '').toLowerCase())) {
    const stage = String(doc.indexStage || 'queued').toLowerCase()
    const labels: Record<string, string> = {
-      queued: '等待处理', extracting: '解析文档', ocr: '扫描件识别', chunking: '切分文本', embedding: '向量化中'
+      queued: '等待处理', extracting: '解析文档', chunking: '切分文本', embedding: '向量化中'
    }
    const progress = Math.max(0, Math.min(99, Number(doc.indexProgress || 0)))
    return { tone: 'pending', label: labels[stage] || '处理中', detail: `${labels[stage] || '正在建立知识索引'} ${progress}%`, progress }
@@ -1,8 +1,5 @@
 export function buildH5Url(base, session) {
  const url = new URL(base)
  // This is deliberately explicit instead of relying on the timing of the
  // H5+ `plus` injection inside the embedded page.
  url.searchParams.set('nativeShell', 'uniapp')
  if (session.token) url.searchParams.set('token', session.token)
  if (session.userId) url.searchParams.set('userId', session.userId)
  if (session.nickname) url.searchParams.set('nickname', session.nickname)