Skip to main content

Composing Layers

The default pipelines provided by ragaxis.ldk are a starting point, not the only way to assemble RAG Axis. Each of the three operational layers (Index, Retrieval, Synthesis) can be configured or replaced independently, as long as the types crossing each boundary still satisfy the contracts described in The Three Layers.

Configuring the Index Layer​

from ragaxis.chunking import SemanticChunker
from ragaxis.embedding import BatchEmbedder

chunker = SemanticChunker(max_chunk_tokens=512, overlap_tokens=64)
embedder = BatchEmbedder(model="text-embedding-3-large")

Swapping SemanticChunker for StructuralChunker or ParentChildChunker changes how documents are split, but every chunker still produces Chunk objects with parent_doc_id and position set, per invariant I1.

Configuring the Retrieval Layer​

from ragaxis.retrieval import HybridRetriever
from ragaxis.reranking import CrossEncoderReranker

retriever = HybridRetriever(store=store, rrf_k=60)
reranker = CrossEncoderReranker(model="cross-encoder/ms-marco-MiniLM-L-6-v2")

The reranker consumes the RetrievalResult produced by the retriever and returns a new RetrievalResult with reordered chunks and normalized scores. If the reranked scores collapse into a narrow range, a ScoreCollapseWarning is attached to the result.

Configuring the Synthesis Layer​

from ragaxis.context import ContextBudget
from ragaxis.generation import PromptTemplate, generate

budget = ContextBudget(max_tokens=4000, reserve_for_response=1000)
template = PromptTemplate.from_file("prompts/answer.j2")

result = generate(query=query, context=context, template=template, budget=budget)

Wiring It Together​

from ragaxis.ldk import build_pipeline

pipeline = build_pipeline(
    chunker=chunker,
    embedder=embedder,
    retriever=retriever,
    reranker=reranker,
    context_budget=budget,
    prompt_template=template,
)

build_pipeline accepts any combination of layer components, as long as each component satisfies the relevant Protocol or produces the expected result type. Components left unspecified fall back to the same defaults used by default_pipeline() in Example: Basic RAG.