The eight stages

Generated by scripts/gen_catalog_doc.py from the component and benchmark registries. Do not edit by hand: run the script.

A RAG pipeline in Lasagna is eight stages behind typed contracts. A component fills one stage, and the stages around it cannot tell which one is there. That is what lets a sweep change one stage and hold the rest fixed.

ingest, chunk, embed, retrieve, enhance the query, rerank, agentic loop, generate

Stage	Runs at	Swap it to change	Built in	Sweep key
Document processor	ingest	how files are parsed into text	4	config_overrides
Chunker	ingest	how documents are split	9	chunkers:
Embedder	query time	what vectors are stored	6	embedders:
Retriever	query time	how candidates are fetched	5	config_overrides
Query enhancer	query time	how the query is rewritten before retrieval	8	query_enhancers:
Reranker	query time	the ordering of retrieved candidates	10	rerankers:
Agentic pipeline	query time	the multi-step retrieval loop	7	config_overrides
Generator	answer time	how the answer is written from the retrieved passages	6	generators:

Every stage can be named in a config_overrides entry with its stage name, for example {name: ircot, agentic_pipeline: ircot}. Five stages also have a list key that sweeps several components at once. A document processor does not sweep: benchmark corpora are loaded as text, so check a processor on your own files.

Registering a component is one decorator. See Write a component.

Document processor

Abstract base class for document processors/extractors.

Base class: BaseDocumentProcessor in lasagna.core.base_components. It runs at ingest. Swap it to change how files are parsed into text.

def extract(file_path: str, **kwargs) -> ProcessedDocument

Function form with @lasagna.document_processor: f(file_path) -> str.

The 4 built-in components are listed in the component catalog.

Chunker

Abstract base class for document chunking strategies.

Base class: BaseChunker in lasagna.core.base_components. It runs at ingest. Swap it to change how documents are split.

def chunk(text: str, **kwargs) -> List[Chunk]

Function form with @lasagna.chunker: f(text) -> list[str].

The 9 built-in components are listed in the component catalog.

Embedder

Abstract base class for embedding models.

Base class: BaseEmbedder in lasagna.core.base_components. It runs at query time. Swap it to change what vectors are stored.

def encode(texts: List[str], **kwargs) -> np.ndarray

Function form with @lasagna.embedder: f(texts) -> one vector per text.

The 6 built-in components are listed in the component catalog.

Retriever

Abstract base class for retrieval strategies.

Base class: BaseRetriever in lasagna.core.base_components. It runs at query time. Swap it to change how candidates are fetched.

def retrieve(query: str, collection: str, top_k: int = 5, **kwargs) -> List[ScoredDocument]

Function form with @lasagna.retriever: f(query, top_k) -> list[str] or documents.

The 5 built-in components are listed in the component catalog.

Query enhancer

Abstract base class for query enhancement strategies.

Base class: BaseQueryEnhancer in lasagna.core.base_components. It runs at query time. Swap it to change how the query is rewritten before retrieval.

def enhance(query: str, **kwargs) -> List[str]

Function form with @lasagna.query_enhancer: f(query) -> list[str].

The 8 built-in components are listed in the component catalog.

Reranker

Abstract base class for reranking models.

Base class: BaseReranker in lasagna.core.base_components. It runs at query time. Swap it to change the ordering of retrieved candidates.

def rerank(query: str, documents: List[ScoredDocument], top_k: int = 5, **kwargs) -> List[ScoredDocument]

Function form with @lasagna.reranker: f(query, text) -> float, or f(query, documents, top_k) -> documents.

The 10 built-in components are listed in the component catalog.

Agentic pipeline

Abstract base class for pluggable agentic RAG pipelines.

Base class: BaseAgenticPipeline in lasagna.core.base_components. It runs at query time. Swap it to change the multi-step retrieval loop.

async def run(input: AgenticPipelineInput) -> AgenticRAGResult

Function form with @lasagna.agentic_pipeline: none: a loop is a class.

The 7 built-in components are listed in the component catalog.

Generator

Abstract base class for answer generation: the last stage.

Base class: BaseGenerator in lasagna.core.base_components. It runs at answer time. Swap it to change how the answer is written from the retrieved passages.

def generate(query: str, contexts: List[str], *, llm: Any = None, instructions: str = '', **kwargs) -> str

Function form with @lasagna.generator: f(query, contexts) -> str (the answer).

The 6 built-in components are listed in the component catalog.