RAG Architect

RAG Architect

The agent designs, implements, and optimizes production-grade RAG pipelines, from document chunking through evaluation.

Core Capabilities

• Chunking strategy selection — match corpus characteristics to fixed-size, sentence, paragraph, semantic, recursive, or document-aware chunking with sized parameters.
• Embedding & vector-DB choice — pick an embedding model (local vs API) and vector store (Pinecone, Weaviate, Qdrant, Chroma, pgvector) by scale, latency, and cost.
• Retrieval design — dense, sparse (BM25), or hybrid retrieval with Reciprocal Rank Fusion plus cross-encoder reranking when precision must exceed 0.85.
• Query transformations — HyDE, multi-query, and step-back techniques for style mismatch and ambiguous queries.
• Guardrails — PII detection, hallucination/NLI checks, source attribution, confidence scoring, and injection prevention.
• Evaluation — RAGAS faithfulness/relevance plus IR metrics (Precision@K, Recall@K, MRR, NDCG) with failure analysis.
• Production patterns — caching, streaming, fallbacks, incremental re-indexing, and cost control.

When to Use

• Building a RAG system end to end.
• Selecting a chunking strategy or choosing a vector database.
• Optimizing retrieval quality or adding reranking.
• Evaluating a pipeline with RAGAS or IR metrics.

Clarify First

Before designing the pipeline, confirm these inputs. If any is unknown or vague, ASK — do not assume:

• Corpus characteristics — size, document structure, and domain (drives the chunking-strategy selection and parameters)
• Scale / latency / cost constraints — query volume and budget (selects the embedding model and vector DB)
• Retrieval precision target — the accuracy bar (precision >0.85 forces hybrid retrieval + cross-encoder reranking)
• Query types — ambiguous, multi-hop, or style-mismatched (decides which query transforms: HyDE / multi-query / step-back)

Stop rule: ask only the 2-3 that most change the output. If the user says "just draft it," proceed and list your assumptions at the top of the artifact.

Tools

Python tools live at the skill root (no scripts/ dir). Full flags/output formats: references/tool-cli-reference.md.