Pipelines

ragway includes five pipeline variants. Set pipeline: in rag.yaml.

naive

The simplest pipeline. Embeds query -> retrieves chunks -> generates answer. Best starting point. Fastest. Lowest cost.

pipeline: naive

hybrid

Combines BM25 keyword search with vector search using Reciprocal Rank Fusion. Better for queries with specific terms, names, IDs, or technical keywords.

pipeline: hybrid
plugins:
  retrieval:
    strategy: hybrid
    hybrid_alpha: 0.5  # 0 = pure BM25, 1 = pure vector

self

Self-correcting pipeline. The LLM grades its own retrieved context and generated answer. Retries if context is irrelevant or answer is unfaithful.

pipeline: self

Returns: { answer, hallucination_score, retries_used }

long_context

Retrieves 20+ chunks, sorts by document position, feeds in order to a long-context model. Best for summarization and questions spanning whole documents.

pipeline: long_context
plugins:
  llm:
    provider: anthropic
    model: claude-opus-4-6  # needs large context window

agentic

LLM uses tools (retrieval, web search, calculator) and loops until it has enough information. Best for multi-step questions.

pipeline: agentic

Returns: { answer, tool_calls_made, iterations_used }