Components

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

55 components ship with the package, across eight stages. Name one in a sweep, or run one with lasagna benchmark run <benchmark> --reranker <name>. The same list prints from lasagna component list.

The last column is what the component needs beyond pip install lasagna-rag. "An LLM" means the run's LLM, set up as in Installation. "Downloads a model" means the weights come from the Hugging Face Hub on first use (see Model licenses). Extras are listed in Installation.

Document processor

4 built in. Swap this stage to change how files are parsed into text. Contract: BaseDocumentProcessor.

Name	What it does	Source	Needs
default	Built-in extraction (pypdf and pdfplumber, OCR when enabled)		Nothing extra
docling	Docling layout analysis, TableFormer tables and OCR, exported as Markdown	arXiv:2408.09869	pip install 'lasagna-rag[docling]'
markitdown	Microsoft MarkItDown: PDF, Office, HTML and more to Markdown	github.com	pip install 'lasagna-rag[markitdown]'
pymupdf4llm	PyMuPDF4LLM: PDF to Markdown with headers and tables (AGPL-3.0)	github.com	pip install 'lasagna-rag[pymupdf4llm]' (AGPL-3.0)

Chunker

9 built in. Swap this stage to change how documents are split. Contract: BaseChunker.

Name	What it does	Source	Needs
contextual	Chunks prefixed with LLM-written document context (Contextual Retrieval)	anthropic.com	An LLM
fixed	Fixed-size windows with overlap, in characters, words or tokens		Nothing extra
hierarchical	Hierarchical parent-child chunker		Nothing extra
markdown_header	Markdown sections with their header path in metadata (LangChain MarkdownHeaderTextSplitter)	github.com	Nothing extra
proposition	Propositions written by an LLM as retrieval units (Dense X Retrieval)	arXiv:2312.06648	An LLM
recursive	Recursive separator splitting: paragraph, line, sentence, word (LangChain RecursiveCharacterTextSplitter)	github.com	Nothing extra
semantic	Semantic chunker		The embedding model
sentence_window	Sliding windows of window_size whole sentences		Nothing extra
structure_aware	Structure-aware chunker		The embedding model

Embedder

6 built in. Swap this stage to change what vectors are stored. Contract: BaseEmbedder.

Name	What it does	Source	Needs
bge_m3	BGE-M3 dense vectors; sparse and multi-vector outputs via FlagEmbedding	arXiv:2402.03216	Downloads BAAI/bge-m3 on first use; [flag-embedding] adds its sparse and multi-vector outputs
cohere	Cohere Embed v3/v4 with search_query/search_document input types	docs.cohere.com	pip install cohere and COHERE_API_KEY (or CO_API_KEY)
default	The configured embedding model		Downloads a model on first use
openai	OpenAI text-embedding-3 small/large, with Matryoshka dimensions	platform.openai.com	openai package and OPENAI_API_KEY
sentence_transformers	Any sentence-transformers model (embedding_model), with its model card's query/document prefixes	arXiv:1908.10084	Downloads a model on first use
voyage	Voyage AI voyage-3 family with query/document input types	docs.voyageai.com	pip install voyageai and VOYAGE_API_KEY

Retriever

5 built in. Swap this stage to change how candidates are fetched. Contract: BaseRetriever.

Name	What it does	Source	Needs
bm25	Okapi BM25 over the session's chunks (bm25_k1, bm25_b)	doi:10.1561/1500000019	Nothing extra
dense	Dense vector search with the configured embedder (the default first stage)	arXiv:2004.04906	Nothing extra
hybrid_convex	Convex combination of normalised dense and BM25 scores (hybrid_alpha)	arXiv:2210.11934	Nothing extra
hybrid_rrf	Dense and BM25 fused by Reciprocal Rank Fusion (rrf_k)	doi:10.1145/1571941.1572114	Nothing extra
splade	SPLADE learned sparse retrieval (splade_model)	arXiv:2109.10086	Downloads a model on first use

Query enhancer

8 built in. Swap this stage to change how the query is rewritten before retrieval. Contract: BaseQueryEnhancer.

Name	What it does	Source	Needs
combined	Diverse LLM reformulations of the query (RAG Fusion style)		An LLM
hyde	HyDE: the query plus a hypothetical answer passage, searched and fused	arXiv:2212.10496	An LLM
multi_query	RAG-Fusion: LLM-generated search queries, fused by reciprocal rank	arXiv:2402.03367	An LLM
query2doc	Query2doc: the query followed by an LLM-written pseudo-document	arXiv:2303.07678	An LLM
query_decomposition	Least-to-most decomposition: the sub-questions a question needs, searched and fused	arXiv:2205.10625	An LLM
query_expansion	LLM paraphrases of the query, searched and fused		An LLM
rewrite	Rewrite-Retrieve-Read: search queries rewritten by an LLM	arXiv:2305.14283	An LLM
step_back	Step-back prompting: the question and a more generic step-back question	arXiv:2310.06117	An LLM

Reranker

10 built in. Swap this stage to change the ordering of retrieved candidates. Contract: BaseReranker.

Name	What it does	Source	Needs
cohere	Cohere Rerank API (pip install cohere, COHERE_API_KEY)		pip install cohere and COHERE_API_KEY (or CO_API_KEY)
colbert	ColBERT late-interaction reranker with token-level MaxSim	arXiv:2004.12832	Downloads a model on first use
combined	Cross-encoder followed by MMR diversity		Downloads a model on first use
cross_encoder	Cross-encoder reranker (the configured reranker_model)	arXiv:1901.04085	Downloads a model on first use
ensemble	Weighted ensemble of cross-encoders		Downloads a model on first use
flashrank	FlashRank ONNX cross-encoders on CPU (pip install flashrank)		pip install flashrank
llm_listwise	RankGPT: LLM permutations over windows sliding from the bottom up	arXiv:2304.09542	An LLM
llm_pointwise	LLM relevance grade per passage on a 0-3 scale (llm_reranker_model)	arXiv:2309.10621	An LLM
mmr	MMR diversity reranking	doi:10.1145/290941.291025	Nothing extra
monot5	monoT5: P(true) from a seq2seq relevance model (monot5_model)	arXiv:2003.06713	Downloads a model on first use

Agentic pipeline

7 built in. Swap this stage to change the multi-step retrieval loop. Contract: BaseAgenticPipeline.

Name	What it does	Source	Needs
adaptive	Adaptive-RAG: a complexity classifier picks no, single-step or multi-step retrieval	arXiv:2403.14403	An LLM
default	Built-in structured RAG graph (13 nodes, CRAG, quality gates)		An LLM
flare	FLARE: retrieve when the tentative next sentence has low-confidence tokens	arXiv:2305.06983	An LLM
ircot	IRCoT: chain-of-thought sentences interleaved with retrieval, each the next query	arXiv:2212.10509	An LLM
legacy_react	Legacy ReAct agent with monolithic search tool	arXiv:2210.03629	An LLM
self_ask	Self-Ask: follow-up questions answered by retrieval until the final answer	arXiv:2210.03350	An LLM
self_rag	Self-RAG (prompted critic): retrieve on demand, critique each passage's answer	arXiv:2310.11511	An LLM

Generator

6 built in. Swap this stage to change how the answer is written from the retrieved passages. Contract: BaseGenerator.

Name	What it does	Source	Needs
chain_of_note	A reading note per passage, then the answer or 'unknown'	Yu et al. 2023, Chain-of-Note: Enhancing Robustness in Retrieval-Augmented LMs	An LLM
chain_of_thought	Step-by-step reasoning, then the final answer	Wei et al. 2022, Chain-of-Thought Prompting Elicits Reasoning in LLMs	An LLM
cited	Answers citing their supporting passages as [n]	Gao et al. 2023, Enabling Large Language Models to Generate Text with Citations (ALCE)	An LLM
closed_book	The question alone, no passages: the no-retrieval baseline		An LLM
concise	One concise sentence from the context (the harness's default prompt)		An LLM
long_context	The concise prompt over all ten retrieved passages, in rank order		An LLM