Note: pipeline

Context

The pipeline module (src/attune_rag/pipeline.py) is a lightweight, LLM-agnostic RAG pipeline. It coordinates four optional components — a corpus, a retriever, a query expander, and a reranker — to retrieve grounding context, assemble a prompt, and optionally call an LLM, all in a single method call.

Content

Two public types form the core of the module:

• RagPipeline — accepts a CorpusProtocol, a RetrieverProtocol, an optional QueryExpander, and an optional LLMReranker at construction time. Its two main methods are:

  • run(query, k, prompt_variant) — retrieves up to k documents, assembles an augmented prompt, and returns a RagResult without calling an LLM.
  • run_and_generate(query, provider, ...) — does the same, then calls the specified LLMProvider and returns (answer_str, RagResult).

• RagResult — the dataclass returned by both methods. Key fields include augmented_prompt, citation (CitationRecord), confidence, fallback_used, elapsed_ms, context, claim_citations, and used_native_citations. When the retriever finds no grounding context, fallback_used is True and the pipeline substitutes FALLBACK_PROMPT_TEMPLATE, which instructs the LLM not to invent APIs, workflow names, or CLI commands.

The prompt_variant parameter selects from the named templates in PROMPT_VARIANTS. The default variant is 'citation'.

Source files

• src/attune_rag/pipeline.py
• src/attune_rag/__init__.py

Tags: pipeline, orchestration, rag, result