Agent skill · data analytics · aradotso
ai-data-engineering-rag-pipeline
Build production-grade local RAG pipelines with BM25 search, hierarchical chunking, and retrieval evaluation from Nahid's hands-on roadmap
Why this skill is useful
Adds executable scripts for building a production-grade RAG pipeline with specific implementations and evaluation methodologies not commonly found in public resources.
What it needs
Requires faiss-cpu, numpy, rank-bm25, sentence-transformers installed locally. About 8k tokens when loaded. Last updated 2026-08-05. 5 stars on the source repository.
What this skill does
AI Data Engineering RAG Pipeline Skill by ara.so — Data Skills collection. This skill helps you build production-grade local RAG (Retrieval-Augmented Generation) pipelines using the architectural patterns and implementations from Nahid Mahmud's AI & Data Engineering Roadmap. The project focuses on baseline search engines, hierarchical chunking strategies, and rigorous evaluation methodologies. What This Project Provides Day 01: Retrieval contract design, governed corpus creation, golden dataset (questions.jsonl), and evaluation frameworks Day 02: BM25 baseline search engine with inverted index, Okapi BM25 ranking, and Recall@10 evaluation Day 03: RAG pipeline with hierarchical chunking (document, section, paragraph levels), deterministic chunk IDs, parent-child metadata linkage, and failure mode analysis Installation & Setup Clone the repository: Install dependencies (each day has its own requirements): Common dependencies across modules: numpy - numerical operations rank-bm25 - BM25 implementation sentence-transformers - embeddings (Day 03+) faiss-cpu - vector search (Day 03+) Project Structure Day 02: BM25 Baseline Search Engine Core Implementation The BM25 baseline provides a deterministic search engine for retrieval evaluation: Recall@10 Evaluation Evaluate retrieval quality against golden dataset: Day 03: Hierarchical Chunking RAG Pipeline Chunk Granularity Strategy Implement multi-level chunking with parent-child metadata: RAG Pipeline with Vector Search Combine BM25 with semantic search: Retrieval Evaluation Framework Evaluate chunking strategies against golden dataset: Configuration Patterns Golden Dataset Format (questions.jsonl) Chunk Metadata Schema Common Patterns Parent-Child Retrieval Retrieve at fine granularity but return parent context: Failure Mode Analysis Identify queries with poor retrieval: Troubleshooting BM25 Returns Low Scores Issue: All BM25 scores are near zero or negative. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills aradotso/ai-data-engineering-rag-pipeline