fix(avatar): ground replies in recognized images

digital-avatar-app/backend/routers/chat.py

@@ -47,6 +47,25 @@ QA_SEMANTIC_THRESHOLD = 0.72
 QA_MATCH_MARGIN = 0.06
 KNOWLEDGE_MIN_SCORE = float(os.getenv("KNOWLEDGE_MIN_SCORE", "0.42"))
 
+_IMAGE_ACCESS_DENIAL_PATTERNS = (
+    re.compile(
+        r"(?:我|目前|暂时|这里|本身|系统)?\s*(?:无法|不能|没法|不支持)\s*"
+        r"(?:直接)?\s*(?:查看|看到|看见|识别|读取|访问|打开|分析|理解)"
+        r"(?:\s*(?:或|、|/)\s*(?:查看|看到|看见|识别|读取|访问|打开|分析|理解))*\s*"
+        r"(?:你(?:发|提供|上传)的|这张|该|当前)?\s*(?:图片|图像|照片|影像|文件)"
+    ),
+    re.compile(
+        r"(?:我|这里|目前|暂时)?\s*(?:看不到|看不见|未看到|没有看到|没收到|未收到)\s*"
+        r"(?:你(?:发|提供|上传)的|这张|该|当前)?\s*(?:图片|图像|照片|影像)"
+    ),
+    re.compile(
+        r"\b(?:i\s+)?(?:can(?:not|'t)|am\s+unable\s+to)\s+(?:directly\s+)?"
+        r"(?:view|see|access|read|analy[sz]e|recogni[sz]e)\s+"
+        r"(?:the\s+|this\s+|your\s+)?(?:image|photo|picture|scan)\b",
+        re.IGNORECASE,
+    ),
+)
+
 _WRITING_SYSTEM_PATTERNS = {
     "han": re.compile(r"[\u3400-\u4dbf\u4e00-\u9fff]"),
     "latin": re.compile(r"[A-Za-z\u00c0-\u024f]"),
@@ -178,6 +197,75 @@ def _image_retrieval_question(question: str, image_contexts: list[dict]) -> str:
     return "\n".join(part for part in parts if part).strip()
 
 
+def _answer_denies_available_image(answer: str) -> bool:
+    """Reject only whole-image access denials, not uncertainty about one field."""
+    value = re.sub(r"\s+", " ", answer or "").strip()
+    return any(pattern.search(value) for pattern in _IMAGE_ACCESS_DENIAL_PATTERNS)
+
+
+def _compact_context_text(value: Any, limit: int) -> str:
+    lines = [re.sub(r"\s+", " ", line).strip() for line in str(value or "").splitlines()]
+    text = "\n".join(line for line in lines if line).strip()
+    return text[:limit].rstrip()
+
+
+def _grounded_image_fallback(question: str, image_contexts: list[dict]) -> str:
+    """Build a safe answer from completed vision data when the chat model contradicts it."""
+    summaries: list[str] = []
+    facts: list[str] = []
+    excerpts: list[str] = []
+    warnings: list[str] = []
+    for context in image_contexts:
+        summary = _compact_context_text(context.get("summary"), 500)
+        if summary:
+            summaries.append(summary)
+        structured = context.get("structuredData") or {}
+        if isinstance(structured, dict):
+            for fact in structured.get("key_facts") or []:
+                value = _compact_context_text(fact, 300)
+                if value:
+                    facts.append(value)
+        extracted = _compact_context_text(context.get("extractedText"), 900)
+        if extracted:
+            excerpts.append(extracted)
+        warning = _compact_context_text(context.get("warning"), 300)
+        if warning:
+            warnings.append(warning)
+
+    summaries = list(dict.fromkeys(summaries))
+    facts = list(dict.fromkeys(facts))[:6]
+    excerpts = list(dict.fromkeys(excerpts))
+    warnings = list(dict.fromkeys(warnings))
+    writing_system = _dominant_writing_system(question)
+
+    if writing_system == "latin":
+        parts = []
+        if summaries:
+            parts.append("From the image, I can confirm: " + " ".join(summaries))
+        if facts:
+            parts.append("Key details:\n" + "\n".join(
+                f"{index}. {fact}" for index, fact in enumerate(facts, 1)
+            ))
+        elif excerpts:
+            parts.append("Visible text:\n" + excerpts[0])
+        if warnings:
+            parts.append("Please note: " + " ".join(warnings))
+        return "\n".join(parts).strip() or "The image is available, but there is not enough clear detail to confirm more."
+
+    parts = []
+    if summaries:
+        parts.append("从这张图中可以确认：" + "；".join(summaries).rstrip("。；") + "。")
+    if facts:
+        parts.append("其中比较明确的信息有：\n" + "\n".join(
+            f"{index}. {fact}" for index, fact in enumerate(facts, 1)
+        ))
+    elif excerpts:
+        parts.append("图中可见的主要文字是：\n" + excerpts[0])
+    if warnings:
+        parts.append("需要注意：" + "；".join(warnings).rstrip("。；") + "。")
+    return "\n".join(parts).strip() or "这张图已经看到了，但目前能确认的清晰信息比较有限。"
+
+
 def _run_billed_vision_call(
     db: Session,
     avatar: Avatar,
@@ -553,8 +641,10 @@ def _build_prompt(
     if image_contexts:
         image_material = json.dumps(image_contexts, ensure_ascii=False, default=str)
         system += (
-            "\n以下是当前会话图片经过视觉识别后得到的资料：\n"
+            "\n当前会话图片已经成功读取并完成内容识别，以下资料就是可直接使用的图片内容：\n"
             f"{image_material}"
+            "\n必须直接依据这些图片内容回答当前问题。禁止声称无法查看、看不到、未收到、无法识别、"
+            "无法读取或不能访问图片，也不要要求对方重新上传；只有资料明确标记读取失败时才可以请对方重发。"
             "\n图片资料可能包含 OCR 错字、模糊内容或用户尚未确认的信息，只能按可见内容谨慎表达。"
             "标准答题对中的事实优先级高于图片资料，知识库事实优先级高于模型推测；发生冲突时遵循更高优先级资料，"
             "并自然提醒对方核对原图。不得声称看到了图片中不存在的内容。"
@@ -815,6 +905,14 @@ def _resolve_reply(
         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(
+            "chat model contradicted ready image context avatar=%s source=%s",
+            avatar.id,
+            usage_source,
+        )
+        answer = _grounded_image_fallback(question, image_contexts)
     result = {
         "answer": answer,
         "source": "qa" if matched else (