project

NovaStock · Goods Ops

FastAPI · PostgreSQL · Function Calling · MCP

Try it live · Opening the demo is faster than reading the write-up.

Open Live Demo →

An internal goods and procurement ops desk: ledger, handling workflow, payment tracking, and file archive — covering the “record → handle → pay” path. The public demo also ships a clear ask-data chain (tool Schema → local data fetch → optional Zhipu tool pick / standard MCP), with an intent-routing case table that locks the triage rules for find-order / stats / get-file.

Live demo · Ask-data lab

Features

  • Source morning brief: payment progress, budget execution, and today’s actions
  • Goods ledger: general-type / no-contract entry points, filters, and status tracking
  • Handling & payments: horizontal stage bar, payment installments, and actual-pay registration
  • Docs center: attachments per document (stable keys when business IDs are empty, so clicks don’t scramble)
  • Ask-data float (self-study): bottom-right “Ask”; rule-based triage by default, optional Zhipu GLM Function Calling; explainable call chain; default year/dept aligned with the dashboard
  • MCP practice: local standard MCP Server (search / stats / attachments), callable from Inspector or Cursor
  • Intent regression: case table keeps “find order ≠ stats ≠ get file”; re-run with a script after rule changes

Tech notes

  • Public demo: static pages + desensitized sample data; production can plug into backend API / PostgreSQL
  • Ask-data: execution layer separated from tool-selection — amounts and lists come from local tools; the model does not hand-calculate
  • Acceptance: ACCEPTANCE.md + smoke scripts + intent-routing-cases.json (a slimmed take on formal intent-regression practice)

Honest note

The ask-data float, mcp-novastock, and demo intent regression are personal self-study builds to explain tool calling, MCP, and regression habits. Full MCP / assistants in a team production environment are collaborative work — interview talk separates those by role.

面向组织内部的货品与采购运营台：台账、办理流程、付款跟踪、资料归档，覆盖「录入 → 办理 → 付款」链路。公开演示站另做了一条可讲清的问数链路（工具 Schema → 本地取数 → 可选智谱选工具 / 标准 MCP），并用意图回归用例表锁住找单 / 统计 / 取文件的分诊规则。

在线演示 · 问数实验室

功能

  • 货源晨报：付款进度、预算执行与今日处置
  • 货品台账：一般类型 / 无合同两支入口，筛选与状态跟踪
  • 办理与付款：横向环节条、付款期次与实付登记
  • 资料中心：按单据查看附件（空业务号用稳定键，避免点选错乱）
  • 问数浮窗（自学）：右下角「问数」；默认规则分诊，可接智谱 GLM Function Calling；调用链可解释；默认年度/部门跟看板对齐
  • MCP 练习：本机标准 MCP Server（检索 / 统计 / 附件），可用 Inspector 或 Cursor 调用
  • 意图回归：用例表约束「找单 ≠ 统计 ≠ 取文件」，改规则后可用脚本复跑

技术要点

  • 公开演示：静态页 + 脱敏演示数据；正式环境可接后端 API / PostgreSQL
  • 问数：执行层与选工具层分离——金额与清单由本地工具返回，模型不手算
  • 验收：ACCEPTANCE.md + 冒烟脚本 + intent-routing-cases.json（借鉴正式系统意图回归思路的精简版）

诚实说明

问数浮窗、mcp-novastock 与演示站意图回归为个人自学实现，用于讲清工具调用、MCP 与回归习惯；团队正式环境中的完整 MCP / 助手为协作产物，面试时按职责分开表述。