Architecture Document
The pipeline stages, the cost lever, the dual cloud/open‑source approach, and what is out of scope.
PDF viewerThe uploaded PDF renders here in the in‑site viewer, page by page. Below is the written summary.
A nine‑stage pipeline
Session context → intent → clarify decision → query rewrite → hybrid retrieval → rerank → evidence gate → constrained generation → verification. Each stage emits a uniform trace row: stage, provenance, model, tokens, cost, confidence.
Figure 1 · pipeline
The cost lever
Only generation and verification are metered. A cloud/open‑source toggle and a budget cap sit in front of them; at the cap the system fails closed to the zero‑cost path rather than degrading silently.
Out of scope, and said so
Ingestion, ticketing and SLA routing are simulated. The multilingual surface is narrowed. The production client, its data and its prompts appear nowhere.