From 853cb21542c861f9915c617f295910b02cc3ab20 Mon Sep 17 00:00:00 2001
From: ericwyuan <ericwyuan@example.com>
Date: Thu, 20 Aug 2026 11:02:26 +0800
Subject: [PATCH] fix: NVIDIA single-frame JSON parsing + Gemini
 timeout/circuit breaker tuning
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

NVIDIA fix:
- Replace parse_vlm_json (requires full schema: global_summary/entities_json/
  frame_details/compute_provider) with lightweight _parse_single_frame_json
  that only extracts per-frame fields (person/action/clothing/etc)
- Root cause: NVIDIA adapter does per-frame analysis returning single-frame
  JSON, but parse_vlm_json rejected it for missing full-schema fields
- Verified: task 297 → 6/6 frames parsed successfully, first SUCCESS

Gemini + circuit breaker tuning:
- Gemini timeout: 30s → 90s (multi-image vision analysis needs more time)
- NVIDIA timeout: 20s → 30s (per-frame API call)
- Circuit breaker threshold: 3 → 5 (less aggressive tripping)
- Circuit breaker cooldown: 600s → 300s (faster recovery)
---
 fam-edge/config/config.yaml                   | 12 +++----
 .../fam_edge/model_adapters/nvidia_adapter.py | 33 ++++++++++++++++---
 2 files changed, 34 insertions(+), 11 deletions(-)

diff --git a/fam-edge/config/config.yaml b/fam-edge/config/config.yaml
index 32a3e08..79f0e68 100644
--- a/fam-edge/config/config.yaml
+++ b/fam-edge/config/config.yaml
@@ -54,11 +54,11 @@ models:
     enabled: true
     model_name: "gemini-flash-latest"   # v1beta 下 gemini-1.5-flash 会 404
     api_key: "${GEMINI_API_KEY}"
-    timeout: 30
+    timeout: 90
     circuit_breaker:
       enabled: true
-      threshold: 3
-      cooldown: 600
+      threshold: 5
+      cooldown: 300
 
   - provider: "nvidia"
     role: "vision"
@@ -66,11 +66,11 @@ models:
     model_name: "meta/llama-3.2-11b-vision-instruct"
     base_url: "https://integrate.api.nvidia.com/v1"
     api_key: "${NVIDIA_API_KEY}"
-    timeout: 20
+    timeout: 30
     circuit_breaker:
       enabled: true
-      threshold: 3
-      cooldown: 600
+      threshold: 5
+      cooldown: 300
 
   # 本地模型：纯文本 qwen2.5:7b，仅参与智能问答，作为 Gemini/NVIDIA 都失败时的兜底
   - provider: "ollama"
diff --git a/fam-edge/src/fam_edge/model_adapters/nvidia_adapter.py b/fam-edge/src/fam_edge/model_adapters/nvidia_adapter.py
index f702b1f..83cd511 100644
--- a/fam-edge/src/fam_edge/model_adapters/nvidia_adapter.py
+++ b/fam-edge/src/fam_edge/model_adapters/nvidia_adapter.py
@@ -10,12 +10,13 @@ SDK: openai (NIM 兼容 OpenAI API 规范)
 """
 import os
 import base64
+import json
+import re
 from typing import Dict, List, Optional
 
 from .base_adapter import BaseModelAdapter
 from .circuit_breaker import CircuitBreaker
 from ..logger import setup_logger
-from ..ai_orchestrator.json_parser import parse_vlm_json, VLMOutputInvalidError
 
 logger = setup_logger('fam-edge.nvidia_adapter')
 
@@ -98,6 +99,30 @@ class NvidiaVisionAdapter(BaseModelAdapter):
         # NVIDIA 单帧无法跨帧综合 global_summary，交由 Edge format_cloud_result 格式化生成
         return {"frame_details": frame_details}
 
+    def _parse_single_frame_json(self, content: str) -> Optional[dict]:
+        """轻量解析单帧 JSON（不要求全 schema，仅提取字段）"""
+        content = content.strip()
+        # 直接解析
+        try:
+            return json.loads(content)
+        except json.JSONDecodeError:
+            pass
+        # 提取 markdown fence
+        fence = re.search(r'```(?:json)?\s*(\{.*?\})\s*```', content, re.DOTALL)
+        if fence:
+            try:
+                return json.loads(fence.group(1))
+            except json.JSONDecodeError:
+                pass
+        # 贪婪匹配最大 {...}
+        brace = re.search(r'\{.*\}', content, re.DOTALL)
+        if brace:
+            try:
+                return json.loads(brace.group(0))
+            except json.JSONDecodeError:
+                pass
+        return None
+
     def _analyze_one_structured(self, path: str, ts: str, idx: int,
                                 known_members: str) -> Optional[Dict]:
         try:
@@ -122,12 +147,10 @@ class NvidiaVisionAdapter(BaseModelAdapter):
             content = resp.choices[0].message.content
             if not content:
                 return None
-            try:
-                data = parse_vlm_json(content)
-            except VLMOutputInvalidError:
+            data = self._parse_single_frame_json(content)
+            if not data:
                 logger.warning(f"NVIDIA 单帧 JSON 解析失败: {content[:120]}")
                 return None
-            # 组装统一字段
             return {
                 "frame_index": idx,
                 "frame_timestamp": str(data.get("frame_timestamp", ts)),