Banking RAG stack seeks production shape

The interesting part is not the stack choice, it is the production discipline around source fidelity, evaluation, and compliance.

• –The scrape-to-Markdown-to-chunks flow is reasonable, but banking content needs structured extraction, page lineage, versioning, and metadata before embedding.
• –Reranking, hybrid search, citations, answer abstention, and regression evals matter more than whether Pinecone or a self-hosted vector DB wins early.
• –PII masking is only one safety layer; the system also needs policy controls, stale-content detection, audit logs, and strict grounding against approved public content.
• –LangChain can help prototype orchestration, but production teams should keep retrieval, prompting, evaluation, and observability modular enough to swap components.