RAG & Embeddings

"Retrieval decides what the model sees; grounding decides whether you can trust what it says." [INFERENCE]

TL;DR

Designs retrieval-augmented generation end-to-end — chunking, embedding-model selection, dense/sparse/hybrid retrieval, reranking, context assembly, and grounding/citation — to reduce hallucination and keep answers traceable to sources. Complements ai-software-architecture (system design) and data-strategy (the corpus). [EXPLICIT]

When to use

• Designing a RAG pipeline (ingest → chunk → embed → retrieve → rerank → generate). [EXPLICIT]
• Choosing an embedding model + similarity metric (cosine/dot/L2). [EXPLICIT]
• Deciding dense vs sparse vs hybrid retrieval for a corpus. [EXPLICIT]
• Reducing hallucination via grounding, citation, and context budgeting. [EXPLICIT]
• NOT for fine-tuning the base model (deferred) or general prompt design (use prompt katas). [INFERENCE]

Sub-capabilities (resource map)

Procedure

1. Chunk by semantic unit with overlap; size to the embedding context + retrieval granularity. [DOC]
2. Embed with a model matched to domain/language; fix the similarity metric. [DOC]
3. Retrieve top-k (hybrid dense+BM25 when recall matters), then rerank for precision. [INFERENCE]
4. Assemble context within a token budget; dedupe; preserve provenance. [INFERENCE]
5. Ground the answer: cite sources, refuse when retrieval is empty/low-confidence. [DOC]

Quality Criteria

• Chunking preserves semantic units with controlled overlap. [INFERENCE]
• Embedding model + metric chosen for domain, not by default. [DOC]
• Retrieval + rerank balance recall and precision. [INFERENCE]
• Answers cite sources; empty/low-confidence retrieval triggers refusal. [DOC]
• Claims evidence-tagged. [EXPLICIT]

Anti-Patterns

• Fixed-size character chunking that splits mid-sentence/table. [INFERENCE]
• Single-vector dense retrieval where exact-term recall matters (use hybrid). [INFERENCE]
• Generating without citation or refusal path (silent hallucination). [DOC]

Contract

• Aceptación: pipeline ingest→chunk→embed→retrieve→rerank→generate con grounding; modelo/metric justificados; citaciones y ruta de refusal. [EXPLICIT]
• Límites: retrieval + generación aumentada; no fine-tuning del modelo base ni diseño de prompts genérico. [EXPLICIT]
• Casos borde: corpus multilingüe → modelo de embedding multilingüe; retrieval vacío → refusal, no inventar. [DOC]
• Supuestos: corpus disponible y un vector store/índice accesible. [SUPUESTO]
• Trade-off: hybrid + rerank suben calidad a cambio de latencia/costo por consulta. [EXPLICIT]

Related Skills

• ai-software-architecture — the system that embeds the RAG component. [EXPLICIT]
• data-strategy — sourcing/governing the corpus. [EXPLICIT]
• notebooklm-research, web-research — gather sources to ingest. [EXPLICIT]

Packet

Capas del packet, cargables bajo demanda (disciplina ICM: una capa por vez, nunca todas juntas): references/ guías de profundidad (cargar UNA por etapa) · knowledge/ cuerpo de conocimiento · prompts/ prompts listos · examples/ salida de ejemplo · agents/ subagentes del packet · assets/ recursos estáticos.