rager 0.1.18

Caching based RAG primitives

rager

Composable RAG primitives with caching baked in.

Install

uv add rager

Getting started

uv run pytest
uv run prek install

For GPU (CUDA/ROCm) torch, add the matching PyTorch index to your own project and install torch from it — those builds aren't on PyPI.

Example

Wire the primitives together yourself — there is no hidden pipeline. This chunks documents, embeds and indexes each chunk, then answers a query from the nearest chunk:

chunker = SemanticChunker()
embedder = SentenceTransformerDenseEmbedder()
index = MemoryDenseIndex()
chunks = MemoryStore()
generator = TransformersGenerator()

for document in documents:
    for chunk in chunker.chunk(document):
       embedding = await embedder.embed(chunk)
       key = await index.add(embedding)
       chunks.set(key, chunk)

query = "Why do cats purr?"
(key,) = await index.similar(await embedder.embed(query), results=1)
answer = await generator.prompt(f"Answer using only the context.\n{chunks.get(key)}\n{query}")

tests/application/ has full dense and hybrid recipes.

API

Every stage is a Protocol with concrete implementations. async methods batch concurrent calls; model-backed methods cache results under .jar/.

Parsers — extract text units from files

• Parser — protocol: units(file) returns text units, id(file) returns the content Hash.
• UnstructuredFileParser — parses any file supported by unstructured.
• PdfFileParser, MarkdownFileParser, CsvFileParser — aliases of UnstructuredFileParser for readable call sites.

Chunkers — split units into chunks

• Chunker — protocol: chunk(unit) -> list[str].
• SemanticChunker — splits on semantic boundaries with a token budget: SemanticChunker(model_name="gpt-3.5-turbo", chunk_size=1000, overlap=0).

Embedders — turn chunks into vectors

• Embedder[E] — protocol: async embed(chunk) -> E.
• SentenceTransformerDenseEmbedder — dense, L2-normalized DenseEmbedding via SentenceTransformers (default all-MiniLM-L6-v2).
• SpladeSparseEmbedder — sparse SparseEmbedding via a SPLADE encoder (default prithivida/Splade_PP_en_v1).

Indexes — store vectors and search by similarity

• Index[E, K] — protocol: async add(embedding) -> key, async remove(key), async similar(embedding, results=100) -> list[key]. Ranks by inner product (equals cosine for L2-normalized vectors). Keys are derived from embedding content, so adding the same vector twice yields one entry.
• MemoryDenseIndex — flat FAISS index; MemoryDenseIndex(dimensions=None) infers width from the first vector unless fixed.
• MemorySparseIndex — in-memory inner-product search over sparse weight maps.
• FileSparseIndex — like MemorySparseIndex, but seals embeddings on disk under .jar/, keeping only keys in memory.

Fusers — merge ranked lists

• Fuser[V] — protocol: fuse(*rankings) -> list[V].
• ReciprocalRankFuser — reciprocal rank fusion with smoothing constant k (default 60).
• BordaCountFuser — Borda count fusion.

Scorers — rerank chunks against a query

• Scorer — protocol: async score(query, chunk) -> float.
• CrossEncoderScorer — cross-encoder reranker (default cross-encoder/ms-marco-MiniLM-L6-v2).

Generators — produce an answer

• Generator — protocol: async prompt(query) -> str.
• TransformersGenerator — local Transformers text-generation model (default HuggingFaceTB/SmolLM2-135M-Instruct, max_new_tokens=512).

Stores — map index keys back to data

• Store[K, V] — protocol: set(key, value), get(key) -> value | None, remove(key), keys().
• MemoryStore[K, V] — in-memory map from key to value (chunk text, embeddings, metadata, ...).
• FileStore[K, V] — like MemoryStore, but seals values on disk under .jar/, keeping only keys and digests in memory.

Types

• Hash — a blake3 hasher; the content ID returned by parsers.
• DenseEmbedding — list[float].
• SparseEmbedding — dict[int, float] mapping token id to weight.