rag-fintech — DeepSeek AI Agent
    Neura MarketNeura Market/DeepSeek
    ChatGPTChatGPTClaudeClaudeGeminiGeminiCursorCursorGrokGrokPerplexityPerplexityDeepSeekDeepSeek
    CoPilotCoPilotStable DiffusionStable DiffusionMidjourneyMidjourney
    View All Directories
    OverviewRulesPromptsMCPsAgentsGamesBlogVideosGuidesCoursesCommunityTrending
    DeepSeekAgentsrag-fintech
    Back to Agents
    rag-fintech

    rag-fintech

    EllenLiu2019 November 30, 2025
    2 copies 0 downloads

    Production-ready RAG engine for medical claim adjudication. FoC (Forest of Clauses) — LLM routes over document structure trees to solve cross-section retrieval failures. Triple-path search · LangGraph Multi-Agent · vLLM Guided Decoding · K8s HPA autoscaling.

    医疗理赔与合规文档 AI 推理引擎

    FoC (Forest of Clauses) 条款森林 — 摒弃传统向量检索“盲人摸象”式的碎片化召回,将文档目录树提升为“一等公民”。通过 LLM Top-down 全局路由与 Vector Bottom-up 碎片检索双引擎并发,彻底解决强层级长文档中“跨章节逻辑对比”的召回死角。

    覆盖文档解析 → 多路检索 → 大模型推理 → 理赔决策全链路的 AI 系统。基于 RAG + GraphRAG + LangGraph Multi-Agent 架构,部署在 Kubernetes 上,支持异构 GPU 算力调度与弹性伸缩。

    <p align="center"> <img src="benchmark/images/pipeline.gif" alt="RAG Pipeline" width="720" /> </p>

    核心亮点

    亮点描述
    FoC 条款森林检索自研 $O(N)$ 单遍状态机解析引擎,动态构建适配主附险的条款森林(5K~10K Tokens),首创由 LLM 基于目录层级执行 Top-down 全局语义路由,并基于 clause_path 实现快速溯源拼装“局部树”,彻底消除跨章节逻辑的召回盲区。
    三路并发检索FoC(结构推理)∥ GraphRAG(实体关系遍历)∥ Vector(语义相似度),三路并发检索 P95 < 1s
    100% 确定性输出深入 vLLM 底层引入 Guided Decoding (FSM 约束),在 Token 采样阶段进行 Logits 拦截,彻底根除高并发下 LLM 输出 JSON 畸形的痛点
    极致上下文压榨针对 FoC 6K+ 长上下文导致的并发灾难,全面开启 vLLM Prefix Caching。实测并发上限从 c≤16 提升至 c=32+,c=28 时 TTFT P99 < 1s
    理赔 Multi-AgentLangGraph 双子图并行(ICD-10 编码 + TNM 分期)→ Human-in-the-Loop 审批 → 时间旅行(Checkpoint fork + replay)
    三级模型路由9B(查询改写,TTFT < 500ms)→ 35B MoE(条款推理,TPOT 24.9ms)→ DeepSeek-Reasoner(理赔决策,0 容错)

    目录

    • 医疗理赔与合规文档 AI 推理引擎
      • 核心亮点
      • 目录
      • 架构概览
      • 模型路由
      • 质量评估
      • 部署指南
        • 外部服务依赖
        • 本地开发
        • K8s 部署
      • 项目结构
      • CI/CD
      • 技术栈
      • 压测报告
        • Qwen3.5-9B 硬件选型对比 (A10 vs L20, 并发 c=4)
        • 长文本并发压测 (Prefix Caching)
      • License

    架构概览

    系统架构:K8s (ACK) 部署,Ingress 统一入口,三级模型路由层 (9B → 35B MoE → DeepSeek-Reasoner),基于 vLLM 原生指标 (num_requests_waiting) 的 HPA 弹性伸缩,HAMi 实现单卡多 Pod 显存隔离。

    存储与检索:PDF 经 LlamaParse 解析后,通过栈式扫描自动构建 ClauseForest,为每个 Chunk 打上 clause_path (GPS定位) 并向量化。持久化到 PostgreSQL (结构数据) + Milvus (向量) + Neo4j (图谱)。检索时三路并发:FoC (LLM 结构路由) ∥ GraphRAG (多跳遍历) ∥ Vector (Dense+Sparse RRF),最终在内存中通过祖先链溯源,拼装出定向修剪的“局部树”喂给 LLM。

    **理赔 Multi-Agent

    Tags

    deepseekgraphraginsurancekuberneteslanggraphmedicalmilvusmulti-agentneo4jqwen

    Comments

    More Agents

    View all
    Klaatcodeagentic-ai

    Klaatcode

    Open-source AI coding agent for the terminal. Claude Code-grade accuracy with smart model routing — uses the right AI model for each task, cutting costs 10x. Supports Claude, GPT, Gemini, DeepSeek & more.

    K
    KlaatAI
    139
    Agentmakeragent

    Agentmaker

    A general-purpose Python framework for building LLM agents and multi-agent systems. "Four lines of code, an agent with memory."

    X
    xinhuangcs
    74
    Api Model Playground Cookbookai-api

    Api Model Playground Cookbook

    Ultimate LLM API Integration Cookbook 2026 for Cursor & AI Agents

    0
    09omerdgn-droid
    150
    Agent Ecologiesagent-framework

    Agent Ecologies

    Ultimate Multi-Agent OS for Autonomous AI NPCs 2026

    I
    israriqbal
    153
    Private Agent

    Private Agent

    PrivateAgent is an open-source Android automation agent built with Flutter. It utilizes the DeepSeek API and native Android Accessibility Services to interpret screen layouts and execute multi-step tasks across any installed application via natural language commands.

    O
    orailnoor
    123
    Loom Novel

    Loom Novel

    把一队分工 Agent 织成一条写小说的流水线,做成桌面客户端;写作指纹让它越写越像你(BYO DeepSeek key,纯本地)。

    W
    WadeZhao23
    184

    Stay up to date

    Get the latest DeepSeek prompts, rules, and resources delivered to your inbox weekly.

    Neura Market LogoNeura Market

    Discover the best AI prompts, plugins, and resources for DeepSeek and more.

    Content Types

    • Rules
    • Prompts
    • MCPs
    • Agents
    • Guides

    Platforms

    • ChatGPT Directory
    • Claude Directory
    • Gemini Directory
    • Cursor Directory
    • Grok Directory
    • Perplexity Directory
    • DeepSeek Directory
    • CoPilot Directory
    • Stable Diffusion Directory
    • Midjourney Directory
    • All Directories

    Resources

    • Blog
    • Documentation
    • Help Center
    • Marketplace

    Legal

    • Privacy Policy
    • Terms of Service

    © 2026 Neura Market. All rights reserved.

    |

    Not affiliated with any AI platform vendors.

    Neura Market

    Custom AI Systems & Services

    Our team of experienced AI builders will help build custom AI systems, workflows, and solutions for your business.

    Request custom work

    Ready-made automations for this

    Workflows from the Neura Market marketplace related to this DeepSeek resource

    • Enhance AI Document Retrieval with OpenAI and Supabase Vector Searchn8n · $14.99 · Related topic
    • AI Chatbot Call Center: Taxi Booking Worker (Production-Ready, Part 5)n8n · $24.99 · Related topic
    • Smarter RAG Agents with Enriched Retrieval and Modular Workflowsn8n · $24.99 · Related topic
    • Build Production-Ready User Authentication with Airtable and JWTn8n · $14.99 · Related topic
    Browse all workflows