scaffold-rag-pipeline

Scaffold RAG Pipeline

Generate a complete retrieval-augmented generation pipeline.

Input: $ARGUMENTS

Steps

1. Load the vector-database skill — read SKILL.md and references/rag-pipeline-patterns.md for templates and anti-patterns.

2. Gather requirements — Extract from $ARGUMENTS or ask:

   • Pipeline name (snake_case, e.g., vendor_rag)
   • Vector store: pgvector, weaviate, or both (two-stage)
   • Embedding model (from skill model table)
   • Chunking strategy (recursive / sentence / paragraph / fixed) + chunk size
   • Reranking: Cohere / cross-encoder / none (must justify "none" for production)
   • Framework: LangChain integration or custom Python
   • Use case (determines retrieval filters and query patterns)

3. Generate pipeline files:

   embedding_service.py

   • EMBEDDING_MODEL constant (not per-call string)
   • EmbeddingService class with embed(text) and embed_batch(texts) async methods
   • Rate-limit and timeout error handling (retry 3× with exponential backoff)
   • try/except on all API calls — no silent failures

   chunking.py

   • Chunking function with strategy from requirements
   • RecursiveCharacterTextSplitter with non-zero overlap
   • Returns List[{"text": str, "metadata": dict}]

   {storage}_retriever.py (pgvector and/or weaviate)

   • Async retrieval function with null guard
   • Structured filter support (city, category, etc.)
   • top_k parameter → retrieves 3× final count for reranking room

   reranker.py (if not "none")

   • Cohere rerank wrapper or cross-encoder
   • Accepts merged candidates from both retrieval stages if two-stage

   pipeline.py

   • Orchestrates: query → embed query → retrieve → rerank → format context
   • Score threshold filter (configurable, default 0.5)
   • Logging of retrieval counts and latency
   • Returns {"results": [...], "context": str, "metadata": {...}}

   tests/test_{pipeline_name}.py

   • Unit test: embedding service returns correct shape
   • Unit test: chunking respects chunk_size and overlap
   • Integration test: retrieve → rerank returns expected structure