Lwindsy / companyagent

# CompanyAgent
English | 中文
---
## English
An enterprise customer-service agent platform: one product, **three interchangeable backends** (Python, Java, .NET) sharing the same API contract, a Vue frontend that switches between them, and a full DevOps pipeline to Azure.
### Architecture
 [code] 
Request flow (`POST /chat`): load memory → recognize intent → retrieve knowledge (query rewrite, hybrid recall, LLM rerank) → route to General / Technical / Billing agent → LLM reply → verify answer → persist memory and update user profile asynchronously.
### Repository layout
| Path | What it is |
|---|---|
| `CompanyAgent/` | Python backend — FastAPI, Anthropic SDK, Redis, ChromaDB, OpenTelemetry, eval suite |
| `CompanyAgentJava/` | Java backend — Java 21, Spring Boot 3.5, Spring AI |
| `CompanyAgentDotnet/` | .NET backend — .NET 8, ASP.NET Core, Polly |
| `CompanyAgentFrontend/` | Vue 3 + Vite frontend with backend switcher |
| `*/skills/` | Business skills as hot-reloadable `SKILL.md` files |
| `deploy/` | Production Docker Compose + Caddy, Kubernetes manifests (kustomize) |
| `infra/terraform/` | Azure infrastructure (VM, network, NSG, monitoring alerts) |
| `.github/workflows/` | CI, CD, and LLM eval gate |
| `docs/` | Deployment runbook, DevOps & observability, lessons learned |
### Key features
- **Multi-agent routing** with hybrid intent recognition and skill-based prompts
- **RAG** with query rewriting, BM25 + vector recall, LLM rerank, caching and circuit breakers
- **Layered memory**: Redis working memory, episodic memory, async user profiles
- **Answer verification** with human-handoff signal and LLM fallback (Claude / DeepSeek)
- **Observability**: Prometheus metrics, Grafana dashboards, alert rules, OpenTelemetry traces (Jaeger)
- **Evaluation**: end-to-end eval set with LLM-as-judge, thresholds and baseline regression checks
### Quick start
Requires Docker and an Anthropic (or Anthropic-compatible) API key.
 [code] 
 [code] 
See each subproject's README for running it standalone.
### CI/CD
| Workflow | Trigger | Does |
|---|---|---|
| `ci.yml` | PR, push to `main` | Ruff + pytest, Terraform validate, kubeconform + compose validation, build all 4 images |
| `deploy.yml` | push to `main` | Run CI → push images to GHCR → SSH deploy to Azure VM → smoke test → **auto-rollback on failure** |
| `eval.yml` | PR touching agents/prompts, daily | Run eval set against the real model; fail on threshold breach or baseline regression |
### Docs
- Deployment runbook
- DevOps & observability
- Lessons learned
- Terraform · Kubernetes
---
## 中文
企业级智能客服 Agent 平台：同一个产品，**三套可互换的后端**（Python、Java、.NET）共用同一套 API 契约，Vue 前端可在三者之间切换，并配有到 Azure 的完整 DevOps 流水线。
### 架构
 [code] 
请求链路（`POST /chat`）：读取记忆 → 意图识别 → 知识检索（查询改写、混合召回、LLM 重排）→ 路由到 General / Technical / Billing Agent → LLM 生成回复 → 回答校验 → 写入记忆并异步更新用户画像。
### 目录结构
| 路径 | 说明 |
|---|---|
| `CompanyAgent/` | Python 后端 —— FastAPI、Anthropic SDK、Redis、ChromaDB、OpenTelemetry、评测集 |
| `CompanyAgentJava/` | Java 后端 —— Java 21、