Ingestion
- Source systems and owners are named
- Ingest jobs are scheduled and monitored
- Failures are visible; silent skip is unacceptable
- Document versions and deletes propagate to the index
Vasiva Checklist
Enterprise RAG Production Checklist
What must be true before retrieval-augmented generation leaves the lab.
Last reviewed: 2026-09-10 · Original Vasiva Framework — not an industry certification
What this is
Enterprise RAG retrieves organization-controlled knowledge, grounds model outputs in that context, and operates under access control, evaluation, and observability requirements typical of production systems.
This page is a Vasiva Framework (or checklist): an original engineering decision aid. It is not an ISO standard, academic instrument, or third-party certification.
Production checklist
Chunking & embeddings
- Chunk strategy matches document types (not one size for all)
- Embedding model and version are pinned and recorded
- Re-embedding plan exists when models change
Retrieval & reranking
- Hybrid retrieval considered where lexical precision matters
- Top-k and filters are explicit
- Reranking (if used) is evaluated, not assumed
- Empty or low-quality retrieval fails closed
Grounding
- Answers must cite retrieved context or refuse
- Unsupported claims are blocked or flagged
- Citation UX makes sources inspectable
Evaluation & confidence
- Gold questions cover real user intents
- Faithfulness / grounding checks run on changes
- Low-confidence paths escalate to humans or refuse
Security & permissions
- Retrieval respects document ACLs
- No cross-tenant leakage in multi-tenant indexes
- Secrets never enter prompts from untrusted docs
Observability & HITL
- Traces show query → retrieval → generation
- Cost and latency budgets are monitored
- Human review exists for high-stakes answers
Deployment
- Index and app deploy are versioned together
- Rollback path is tested
- Kill switch disables generation without deleting data
Scoring method
Use as a binary readiness checklist per item (pass / fail / not applicable). Do not ship production RAG with open fails on security, ACL isolation, grounding, or rollback. This is a Vasiva Checklist — not a published industry benchmark.
Direct answers
When should an enterprise use RAG?
When answers must be grounded in organization-controlled documents or data that change over time, and when those sources can be permissioned, evaluated, and observed in production.
How do you reduce hallucinations in enterprise RAG?
Retrieve relevant permissioned context, require grounding or refusal, evaluate faithfulness on a fixed set, and escalate low-confidence answers to humans. Do not let the model grade its own correctness as the only control.
Related services
Related frameworks
If you want a structured review of a live initiative against this framework, talk to the engineers who ship Vasiva products.
Assess Your RAG Architecture