rag-architect

---
name: rag-architect
description: Use when designing RAG pipelines, choosing chunking strategies, selecting embedding models, or evaluating retrieval quality against ground truth.
category: Engineering
version: 1.0.0
tools: []
---

# RAG Architect

Design, tune, and evaluate production RAG pipelines with three deterministic tools.

## Hard Rules

1. **Never present model names or vendor prices as current facts.** Embedding models and vector-DB pricing change monthly. Recommend a tier, name a candidate, and tell the user to verify against live pricing.
2. **Every design ends with evaluation.** A RAG design without metric numbers is a hypothesis, not a deliverable.
3. **Chunking is corpus-driven.** Run analysis on the real documents before choosing a strategy.

## Embedding Model Tiers

| Tier | Current Examples (verify) | When |
|---|---|---|
| Fast / self-hosted | `all-MiniLM-L6-v2`, `bge-small` | Cost-sensitive, small scale, real-time |
| Balanced open | `all-mpnet-base-v2`, `bge-large`, `e5-large` | Quality without API dependency |
| Quality API | `text-embedding-3-large`, `voyage-3-large` | Accuracy-priority general retrieval |
| Code | `voyage-code-3`, CodeBERT | Code search corpora |

Always note: "verify current pricing/availability before use."

## Workflow

### 1. Corpus Analysis

Analyze document types, size, and structure. Determine optimal chunking strategy (fixed-size, semantic, sliding window, recursive).

### 2. Design the Pipeline

Define requirements: document count, query patterns, latency SLA, budget, accuracy priority. Output pipeline architecture, chunking config, embedding choice, vector DB (Pinecone, Weaviate, Qdrant, Chroma, pgvector), reranking strategy.

### 3. Evaluate Retrieval Quality

Prepare query set and ground-truth labels. Measure:
- Precision@k — how many top-k results are relevant
- Recall@k — coverage of all relevant docs in top-k
- MRR — mean reciprocal rank
- NDCG@k — normalized discounted cumulative gain

### 4. Iterate

If metrics fall below targets (typical: precision@5 ≥ 0.8, recall@10 ≥ 0.85):
- Inspect poor-precision examples; adjust chunking or embedding
- Inspect poor-recall examples; use hybrid retrieval or expand corpus
- Re-evaluate; repeat

## Vector Databases

- **Managed**: Pinecone, Weaviate Cloud (ease, scale)
- **Self-hosted**: Qdrant, Chroma, pgvector (control, cost)
- Selection criteria: scale, filtering, existing Postgres, cost, uptime SLA

## Evaluation Metrics

Precision@k and recall@k are standard; NDCG handles ranking quality. Always measure before optimizing.

## References

For chunking strategy trade-offs, embedding model benchmarks, and metric definitions, search the Knowledge Base.