RAG Patterns — Retrieval-Augmented Generation

Operating contract

Inputs

Outputs

• Produce: retrieval architecture, ingestion/chunking policy, index and filter contract, grounded-answer schema, and evaluation report.

Capability and permission boundaries

Default to read-only analysis. Read only scoped records; redact secrets and regulated data. Writes, execution, network calls, production configuration, customer communication, billing changes, and delegation require explicit authority and an identified owner. Never widen tenant, time-window, or system scope implicitly.

Degraded mode

When required telemetry, evidence, execution, network access, or write authority is unavailable, return a partial result with each unassessed item labelled, preserve the safest existing state, and state the evidence or approval needed to continue. Never convert missing evidence into a pass.

Decision rules

Anti-Patterns

• Treating absent evidence as success. Fix: mark the check unassessed and name the missing source.
• Expanding one tenant or workflow to all tenants. Fix: enforce supplied scope at every query and action.
• Performing a production write during analysis. Fix: emit a reviewed change plan until authority is explicit.
• Reporting a metric without population, window, or source. Fix: attach all three.
• Hiding a failed threshold inside an average. Fix: report failure slices and the remediation owner.

Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com, +256 784 464178.

Use When

• Use when building features that answer questions from private data, documents, policies, or time-sensitive information — RAG architecture, chunking strategies, hybrid search, re-ranking, vector databases, evaluation, agentic RAG, multimodal RAG...

Evidence Produced

References

• Use the references/ directory for deep detail after reading the core workflow below.
• references/curated-corpus-worked-example.md: small curated corpus with BM25 ranking, per-domain score floors, calibrated abstention and a graded, fingerprinted golden set.

Overview

RAG solves the core LLM limitation: they only know what they were trained on. Use RAG to inject private data (invoices, menus, policies, reports) into every AI response.

Core principle: RAG = look up a database + LLM synthesises the results. The LLM never needs to "know" your data.

When to Use RAG

RAG vs Fine-Tuning

Default: start with RAG. Fine-tune only when RAG + prompt engineering cannot deliver the required tone or vocabulary.

Additional Guidance

Guidance is split across two reference files so this entrypoint stays compact.

references/skill-deep-dive.md — architecture, chunking, retrieval, schema:

• Pipeline Architecture
• Chunking Strategies
• Embedding Model Selection
• Vector Database Selection
• Retrieval Algorithms
• Re-Ranking
• Full RAG Query Algorithm
• Query Rewriting (Multi-Turn)
• RAG Schema (Multi-Tenant)
• Evaluation Framework
• Production Patterns
• Agentic RAG
• Multimodal RAG, Edge Cases, Cost Optimisation, Sources

references/production-rag.md — the progression from draft to production and the gates before shipping:

• RAG Maturity Model — Naive → Advanced → Modular
• Query Transformation — HyDE, Multi-Query, Step-Back
• Contextual Compression
• Self-RAG
• RAGAS Evaluation — 4 metrics with production thresholds
• Embedding Pipeline — batching, upserts, re-embed triggers, $/1M-token table
• Cost Management Decision Tree — concrete dollar figures per branch
• Failure Mode Playbook — empty, irrelevant, hallucinated, stale
• Gates Before Shipping

references/iterative-retrieval.md — load when a context-starved subagent must explore a codebase or workspace before working: bounded DISPATCH-EVALUATE-REFINE-LOOP (max 3 cycles), relevance banding, explicit gap lists, and terminology learning.

Load the production file when building a RAG system that has to pass evaluation gates, survive multi-tenant review, or hit a cost budget under load.

When answers depend on structured records about a specific customer, order or account (not documents), or when embedding, prompt and model versions must be managed like an ML system, load python-ml-predictive reference ml-system-architecture-fti.md (feature/training/inference pipelines, entity-keyed retrieval, version coupling).

Multi-Tenant Addendum

This skill describes RAG patterns in general. When the RAG feature ships inside a multi-tenant SaaS, the production answer is ai-rag-multi-tenant — per-tenant ingestion pipelines, vector store partitioning, tier-specific chunking and embedding models, defence-in-depth retrieval security, and citation grounding tied to live sources.

Cross-references:

• ai-rag-multi-tenant — multi-tenant RAG end-to-end.
• ai-tenant-isolation-patterns — vector-store partitioning tradeoffs and data-bleed tests.
• ai-on-saas-architecture — KB service as a control-plane service.
• ai-hallucination-slo-and-grounding — citation grounding + faithfulness SLO.
• ai-model-gateway — gateway-mediated retrieval calls.
• saas-tenant-data-portability-and-erasure — KB erasure cascade for embeddings.
• Catalog metadata for AI — load when retrieval or an agent draws on data-catalog or data-contract metadata: classification-first filtering, ODCS context block, provenance.

Consolidated Child References

• Load references/routing.md to map retired AI child skill slugs to their reference modules.