RAG Architecture Designer

Purpose

Design a Retrieval-Augmented Generation (RAG) pipeline architecture: ingestion, indexing, retrieval, generation, and evaluation components.

Input: Use case, document types, query patterns, quality/latency requirements
Output: RAG Architecture document with pipeline diagram, component specs, and technology recommendations
References: See references.md for domain cheat sheets. Examples: See examples.md for worked input/output.

Workflow

Step 1: Define RAG requirements

Step 2: Design ingestion pipeline

Source → Extract → Clean → Chunk → Embed → Index

Document per stage: format handlers, metadata extraction, update frequency.

Step 3: Design retrieval pipeline

Query → Preprocess → Retrieve → Rerank → Context assembly → Generate

Specify retriever type: dense, sparse, hybrid.

Step 4: Select vector store and index

Consider: pgvector, Pinecone, Weaviate, OpenSearch, managed vs self-hosted.

Step 5: Plan evaluation loop

• Offline: golden Q&A set, RAGAS metrics
• Online: user feedback, citation click-through
• Regression on corpus updates

Step 6: Validate

Run Validation checklist.

Decision Rules

Validation

• Query types and corpus characteristics documented
• Ingestion pipeline end-to-end defined
• Retrieval pipeline with retriever type specified
• Vector store recommendation with trade-offs
• Context assembly linked to context-engineering budget
• Eval plan with offline and online metrics
• Latency budget allocated per pipeline stage
• Citation strategy defined for customer-facing use

Anti-patterns

• RAG by default — applying RAG without retrieval need.
• Chunk and pray — no chunking strategy or eval.
• Single retriever — no hybrid when queries vary widely.
• No freshness plan — stale index without update pipeline.
• Missing citations — generated answers without source attribution.

Best Practices

• Start with smallest pipeline that passes eval bar.
• Separate ingestion from query serving for independent scaling.
• Version indexes for rollback on bad ingests.
• Align with ai/context-engineering for assembly.
• Document RAG-specific ADRs.

Output Structure

# RAG Architecture: [System Name]

## Requirements
| Dimension | Target |
|-----------|--------|

## Ingestion Pipeline
[Stage descriptions]

## Retrieval Pipeline
[Stage descriptions]

## Technology Stack
| Component | Choice | Rationale |
|-----------|--------|-----------|

## Evaluation
| Metric | Target | Method |
|--------|--------|--------|

## Diagram
```mermaid
[pipeline diagram]

---

# Next Skills

| Outcome | Recommended Skill |
|---------|-------------------|
| Chunking strategy | `rag/chunking-strategy-advisor` |
| Embedding selection | `rag/embedding-strategy-advisor` |
| Optimize retrieval | `rag/retriever-optimizer` |
| Hybrid search | `rag/hybrid-search-advisor` |
| Context budget | `ai/context-engineering` |