RAG Expert
Overview
Retrieval Quality determines Generation Quality.
This skill provides the architecture and code patterns for building production-grade Retrieval-Augmented Generation systems. It enforces "Garbage In, Garbage Out" discipline.
[!IMPORTANT] Key Principle: Never trust raw retrieval. Always rerank.
Quick Example (30 sec)
python
# 1. Semantic Chunking (Context is King) # See templates/semantic_chunking.py for full implementation chunks = semantic_chunking(documents) # 2. Hybrid Search (Keywords + Vectors) # See templates/hybrid_search.py for full implementation retriever = setup_hybrid_search(chunks, embeddings, Chroma) # 3. Rerank & generate # Always rerank results before sending to LLM final_docs = reranker.compress_documents(retrieved_docs, query)
Core Principles
1. Chunking Strategy
Don't slice by character count. Use semantic boundaries (sentences, paragraphs). Bad chunks = context loss = hallucination.
- •Reference:
templates/semantic_chunking.py
2. Hybrid Search
Pure vector search is insufficient. Vectors miss specific keywords (IDs, names). BM25 misses concepts. Use both.
- •Reference:
templates/hybrid_search.py
3. Evaluation First
Measure retrieval separately. If you improved the prompt but reduced retrieval precision, you failed.
- •Reference:
templates/retrieval_eval.py
Excusa vs Realidad (Guardrails)
| Excusa del Agente | Realidad (Lo que debes hacer) |
|---|---|
| "Solo usaré embeddings de OpenAI por defecto" | Selecciona el modelo correcto. Revisa references/embedding_models.md para elegir según el caso (código, multilingüe, etc.). |
| "Un chunk size de 1000 es estándar" | Depende del contenido. Usa chunking semántico, no un número arbitrario. |
| "No necesito reranking, mi vector store es bueno" | Mentira. El reranking siempre mejora el Precision@k. Hazlo. |
| "El usuario no pidió evaluación" | Implementa métricas básicas. Sin Recall@k vuelas a ciegas. |
Architecture Patterns
Hierarchical Retrieval
Index summaries for fast search, retrieve full chunks for generation.
- •Summary Index -> Search
- •Full Doc Store -> Retrieve
Contextual Reranking
Use an LLM or Cross-Encoder to rescore top-K results.
- •Input: Query + Top 50 docs
- •Output: Top 5 relevant docs (re-ordered)