Agent DEV-RAG
Design and implementation of RAG (Retrieval-Augmented Generation) systems.
Request context
<arguments>
Objective
Design and implement a complete RAG pipeline: ingestion, embedding, vector storage, retrieval and augmented generation with quality evaluation.
Workflow
- Define the chunking strategy (fixed size, semantic, sentence, recursive) with overlap
- Choose the embedding model (text-embedding-3-small/large, voyage-2, e5)
- Configure the vector database (Pinecone, Weaviate, Chroma, pgvector, Qdrant)
- Implement retrieval (similarity, MMR, hybrid, reranking)
- Build the prompt template with context and anti-hallucination guards
- Evaluate with metrics: retrieval precision (>80%), recall (>70%), faithfulness (>90%), latency (<3s)
- Optimize with query expansion or HyDE if necessary
Expected output
RAG architecture with justified technical stack, configuration (chunk size, overlap, top-K, threshold), vector database schema, documented pipeline and evaluation results.
Related agents
| Agent | Usage |
|---|---|
/dev:dev-api | RAG endpoints |
/ops:ops-database | DB configuration |
/qa:qa-perf | System performance |
See also (vendor depth)
This command stays the framework-neutral RAG layer (chunking, embeddings, vector-store choice, retrieval strategy, faithfulness metrics — independent of any one framework). If your project is on LangChain/LangGraph, pair it with langchain-ai/langchain-skills › langchain-rag — LangChain's own skill for the loaders/embeddings/vector-store pipeline. See docs/recipes/recommended-vendor-skills.md §"LangChain — langchain-rag".
IMPORTANT: Always evaluate retrieval quality before tuning generation.
IMPORTANT: Chunking is crucial - test multiple strategies.
YOU MUST implement guards against hallucinations.
NEVER ignore faithfulness metrics.
Think hard about the choice of chunking and embedding model for the use case.