Text-only e-commerce customer service

This example shows a small Chinese-language LangGraph customer-service workflow with LoongSuite automatic instrumentation. It accepts text only: there is no image input, file upload, browser UI, or multimodal model.

It targets LangGraph 1.2+ and uses LangChain 1.x create_agent, the supported LangGraph-backed agent API.

The workflow is:

Chinese customer question
  -> intent_router
  -> presales_agent | aftersales_agent | clarify
  -> response_review
  -> final Chinese text response

The router, specialist prompts, synthetic data, tool results, CLI, and customer responses use Simplified Chinese. The specialist Agent display names are 售前服务 and 售后服务, so their observable spans are named invoke_agent 售前服务 and invoke_agent 售后服务. Stable Python identifiers, graph node names, tool names, and route values remain in English. The two specialist branches use separate prompts and tools. All products, orders, policies, and tool results are fictional fixtures in tools.py.

Install

Python 3.10 or later is required.

python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt

Configure a model

The default uses the DashScope OpenAI-compatible endpoint and qwen-plus:

export DASHSCOPE_API_KEY="your-api-key"

Override the compatible endpoint or model when needed:

export OPENAI_API_KEY="your-api-key"
export OPENAI_BASE_URL="https://your-compatible-endpoint/v1"
export MODEL_NAME="your-model"

DASHSCOPE_API_KEY takes precedence over OPENAI_API_KEY. Do not commit keys to this directory.

Run with LoongSuite

Console spans are convenient for a local smoke test:

export OTEL_SERVICE_NAME="ecommerce-customer-service"
export OTEL_TRACES_EXPORTER="console"
export OTEL_METRICS_EXPORTER="none"
export OTEL_LOGS_EXPORTER="none"
export OTEL_SEMCONV_STABILITY_OPT_IN="gen_ai_latest_experimental"

loongsuite-instrument python app.py \
  --question "云步通勤鞋适合日常步行吗？42 码有货吗？"

After-sales example:

loongsuite-instrument python app.py \
  --question "订单 DEMO-1001 昨天签收，鞋底有问题，应该怎么处理？"

Omit --question to start a simple interactive Chinese text loop.

For a detailed scenario walkthrough and a ready-to-use training script, see README.zh-CN.md.

With LangChain and LangGraph instrumentation enabled, a specialist request is expected to include the router LLM, the selected Agent, ReAct Step, Tool/LLM children, and the final reviewer LLM. Only the chosen specialist branch runs. The outer graph nodes remain presales_agent and aftersales_agent, while the inner Agent spans use the Chinese display names above.

Test

The offline tests do not call an external model:

python -m pip install -r requirements-dev.txt
python -m pytest tests -q

They cover routing, separate specialist branches, synthetic tool results, missing orders, fail-open behavior, and concurrent state isolation.

Privacy boundary

This example is deliberately generic. P-DEMO-* products and DEMO-* orders are synthetic and are not derived from a real store or customer environment.