Skip to Content
TemplatesBasic

basic template

Vector-only RAG pipeline for developers building their first production-grade document Q&A system.

What you get

project/
├── src/
│   ├── pipeline.py       # Main RAG pipeline — ingest and query
│   └── config.py         # Pipeline configuration
├── eval/
│   ├── golden_set.json   # Evaluation dataset (add your Q&A pairs)
│   └── config.yaml       # Evaluation thresholds and CI gate settings
├── pyproject.toml        # Python dependencies and RAG-Forge config
└── README.md

Default configuration

The basic template uses recursive chunking with dense vector retrieval backed by Qdrant. There is no BM25 or reranker — retrieval is purely embedding-based.

# pyproject.toml [tool.rag-forge] section
[tool.rag-forge]
template = "basic"
chunk_strategy = "recursive"
chunk_size = 512
overlap_ratio = 0.1
vector_db = "qdrant"
embedding_model = "BAAI/bge-m3"

Evaluation is gated on faithfulness >= 0.85 and tracks context_relevance, faithfulness, and answer_relevance.

  1. Add your documents to ./docs and run rag-forge index --source ./docs to build the vector index.
  2. Add representative question-answer pairs to eval/golden_set.json, then run rag-forge audit --golden-set eval/golden_set.json to measure baseline quality.
  3. Swap the embedding model in pyproject.toml to match the language or domain of your documents (e.g., a multilingual model for non-English content).

When to upgrade

Move to the hybrid template when vector-only retrieval misses keyword-heavy queries — for example, exact product codes, technical identifiers, or any domain where BM25 recall complements dense retrieval.

n8nHybrid