Adaptive-RAG 电商客服

A Chinese-language e-commerce customer-service system from gbh1234, built to fix answers that lack evidence in multi-condition policy questions. FastAPI, LangGraph and React drive adaptive retrieval routing, multi-round evidence gathering, fidelity checks and a chat UI with policy citations and order tools.

Jev replaces the intent-classification node: it uses the structured choice interface to return labels, confidence and class probabilities for 14 predefined business intents, and responses are validated with Pydantic before a rule overlay corrects confusable cases. Jev is called over HTTP with no special SDK.

Edit the local .env: set INTENT_BACKEND=jev, JEV_API_KEY, JEV_ENDPOINT (https://api.typesafe.ai/v1/systemone by default in the sample) and JEV_MODEL, configured as jev-1.13.0; timeout, retries and the overlay flag are also configurable. Restart the backend, and revert with INTENT_BACKEND=deepseek.

Missing keys, failed requests or invalid responses fall back to the original IntentClassifier, which prefers the configured DeepSeek model and falls back to rule-based classification offline. Retryable errors use exponential backoff, and JEV_MAX_RETRIES counts maximum attempts.

Python 3.11+ and Node.js 20.19+ or 22.12+ are required. Create a virtual environment, install with pip install -e ".[dev]", copy .env.example to .env, run the demo seed script, start uvicorn on port 8000, and run the frontend with npm ci and npm run dev on port 5173.

No. The README marks it suitable for local demos, learning and engineering validation: orders and logistics are demo data, the API does not verify the user_id it receives, pending actions live in server process memory for 15 minutes, and public deployment needs authentication, TLS and persistence work.

Yes, eval.compare_intent --backend both runs on the same samples and overlay settings and reports accuracy, P50/P95 latency, fallback counts and overlay counts. The README warns the current set is a development set, accuracies include fallback samples, and historical DeepSeek numbers are not Jev scores.