feat: 重构知识库系统，移除Hermes集成，增强RAG和同步功能

主要变更:
- 移除Hermes智能体及相关回调服务
- 新增知识库RAG、同步、调度、规范化和索引任务服务
- 重构orchestrator服务，增强运行时聊天功能
- 更新前端聊天、政策制度、设置等页面样式和逻辑
- 更新expense_claims和document_intelligence服务
- 删除llm_wiki相关服务和测试文件
- 更新docker-compose配置和启动脚本

server/src/app/services/document_intelligence.py

@@ -9,8 +9,6 @@ from typing import Any
 from pydantic import BaseModel, Field, ValidationError
 from sqlalchemy.orm import Session
 
-from app.services.runtime_chat import RuntimeChatService
-
 
 @dataclass(frozen=True, slots=True)
 class DocumentField:
@@ -198,7 +196,7 @@ MERCHANT_PATTERNS = (
 
 class DocumentIntelligenceService:
     def __init__(self, db: Session | None = None) -> None:
-        self.runtime_chat_service = RuntimeChatService(db) if db is not None else None
+        self.db = db
 
     def build_document_insight(
         self,
@@ -254,95 +252,6 @@ class DocumentIntelligenceService:
         rule_insight: DocumentInsight,
         fields: tuple[DocumentField, ...],
     ) -> tuple[str, LlmDocumentClassification] | None:
-        if self.runtime_chat_service is None:
-            return None
-
-        trimmed_text = text.strip()
-        if not trimmed_text and not summary.strip():
-            return None
-
-        facts = {
-            "filename": filename,
-            "summary": summary[:300],
-            "ocr_text_excerpt": trimmed_text[:2000],
-            "rule_candidate": {
-                "document_type": rule_insight.document_type,
-                "document_type_label": rule_insight.document_type_label,
-                "scene_code": rule_insight.scene_code,
-                "scene_label": rule_insight.scene_label,
-                "expense_type": rule_insight.expense_type,
-                "confidence": round(rule_insight.classification_confidence, 2),
-                "evidence": list(rule_insight.evidence),
-            },
-            "extracted_fields": [
-                {"key": field.key, "label": field.label, "value": field.value}
-                for field in fields
-            ],
-            "allowed_document_types": list(SUPPORTED_DOCUMENT_TYPES),
-        }
-
-        system_prompt = (
-            "你是企业报销票据识别复核器。"
-            "你的任务不是 OCR，而是在已有 OCR 文本和票据预览基础上判断票据类型，并尽量复核关键字段。"
-            "只输出 JSON 对象，不要输出 Markdown、解释或代码块。"
-            "document_type 只能是："
-            f"{', '.join(SUPPORTED_DOCUMENT_TYPES)}。"
-            "如果证据不足，返回 other。"
-            "严禁编造 OCR 中不存在的商户、酒店、航司、路线或金额。"
-            "如果 OCR 出现冲突碎片，应优先依据票据主体信息，而不是单个噪声词。"
-            "例如滴滴行程单/网约车发票，即使 OCR 混入酒店名称，也不能直接判成酒店票据。"
-            "如果能从 OCR 或图片中明确确认字段，可在 fields 中返回。"
-            "fields 只允许包含 key, label, value，key 只能是 amount, date, merchant_name, invoice_number, "
-            "invoice_code, trip_no, route。无法确认就不要返回该字段。"
-            "输出字段：document_type, scene_code, scene_label, expense_type, confidence, evidence, fields。"
-        )
-        user_prompt = (
-            "请根据以下票据事实给出最终分类 JSON：\n"
-            f"{json.dumps(facts, ensure_ascii=False, indent=2)}\n\n"
-            "示例输出：\n"
-            "{\n"
-            '  "document_type": "taxi_receipt",\n'
-            '  "scene_code": "transport",\n'
-            '  "scene_label": "交通票据",\n'
-            '  "expense_type": "transport",\n'
-            '  "confidence": 0.86,\n'
-            '  "evidence": ["OCR 中出现 滴滴出行、订单号、上车/下车 等交通特征"],\n'
-            '  "fields": [{"key": "amount", "label": "金额", "value": "32.5"}]\n'
-            "}"
-        )
-
-        if preview_data_url:
-            response_text = self.runtime_chat_service.complete(
-                [
-                    {"role": "system", "content": system_prompt},
-                    {
-                        "role": "user",
-                        "content": [
-                            {"type": "text", "text": user_prompt},
-                            {"type": "image_url", "image_url": {"url": preview_data_url}},
-                        ],
-                    },
-                ],
-                slot_priority=("vlm",),
-                max_tokens=320,
-                temperature=0.0,
-            )
-            parsed = self._parse_llm_payload(response_text)
-            if parsed is not None:
-                return "llm_vision", parsed
-
-        response_text = self.runtime_chat_service.complete(
-            [
-                {"role": "system", "content": system_prompt},
-                {"role": "user", "content": user_prompt},
-            ],
-            slot_priority=("main", "backup"),
-            max_tokens=320,
-            temperature=0.0,
-        )
-        parsed = self._parse_llm_payload(response_text)
-        if parsed is not None:
-            return "llm_text", parsed
         return None
 
     @staticmethod