About this project

The repository implements a customer-operations control plane where an LLM interprets requests and proposes semantic intents, but cannot directly call mutation tools. The server handles authentication, tenant and customer scope, target resolution, business validation, policy, confirmation, post-confirmation revalidation, typed execution, idempotency, and database constraints. Supported support tasks include order and ticket lookup, controlled refunds and cancellations, ticket creation, and human escalation. Knowledge answers use Qdrant-based dense/BM25 hybrid retrieval, excerpt and citation validation, and bounded answers that abstain when evidence is insufficient. RAG and memory are explicitly context-only and cannot grant execution authority. Workflows are built with LangGraph and PostgreSQL checkpoints, allowing suspension, resumption, replacement, and recovery across restarts. FastAPI and Pydantic provide the API; SQLAlchemy and Alembic handle persistence; OpenTelemetry and Jaeger support tracing. The React/TypeScript operator console presents a read-only request-to-outcome investigation view, including evidence, proposals, decisions, authority state, execution, lifecycle, and replay results while omitting raw prompts, secrets, and hidden reasoning. The project ships Docker Compose setup, seeded data, a bundled knowledge base, backend and frontend tests, safety and resilience evaluations, and Playwright journeys. The README documents reported evaluation and release-check results, but describes the system as a production-oriented reference implementation rather than a managed service or compliance certification, with deployment-specific security and capacity concerns remaining outside its scope.