Components¶

The pluggable adapters, grouped by pipeline stage. Each family has a port in Contract (core); the classes below are the reference adapters that declare it. Heavy adapters import their backend lazily and require an extra (noted per class) — installing the core package never pulls them in.

Serializing¶

`TemplateSerializer`	`"Name: {name}"`; `column_mapping` maps this source's columns onto the template variables.
`FieldwiseSerializer`
`WholeRowSerializer`	Package default when `Source(serializer=None)`.
`default_serializer`

Embedding¶

`HashedNGramEmbedder`	Dependency-free reference embedder.
`SentenceTransformerEmbedder`	Semantic embedder over a `sentence-transformers` checkpoint (extra: `[sentence-transformers]`).

Indexing¶

`NumpyFlatIndex`	Dependency-free reference index spec (brute-force / flat search).
`NumpySearchableIndex`	Immutable artifact built by `NumpyFlatIndex` — exhaustive (flat) nearest-neighbour search by inner product (which equals cosine for the L2-normalized vectors the reference embedder produces).
`FaissFlatIndex`	FAISS-backed reference index spec — exact (flat) inner-product search (extra: `[faiss]`).
`FaissSearchableIndex`	Immutable artifact built by `FaissFlatIndex` — exact (flat) inner-product nearest-neighbour search over a `faiss.IndexFlatIP`.

Blocking¶

`DenseBlocker`	Dense-blocking spec.
`DenseBlockingIndex`	Immutable artifact built by `DenseBlocker` for one reference set.

Filtering¶

SimilarityThresholdFilter

Keeps pairs whose similarity_score >= threshold.

Matching¶

`ThresholdMatcher`	Dependency-free reference matcher; gates on the carried similarity.
`LangChainMatcher`	LLM matcher (extra: `[langchain]`).
`RetryPolicy`

Clustering¶

`ConnectedComponentsClusterer`	Reference `Clusterer` adapter; delegates to `denselinkage.clustering.connected_components()`.
`connected_components`	Connected-components clustering: transitively close the matched pairs in `result` and label each record with its component id.

Mining¶

mine_hard_negatives

The hardest negatives among candidates: scored pairs not in gold, ordered by descending similarity (ties broken by id for determinism).