API Reference

On this page

Public classes
Common interface (from BaseEmbedder)
Data classes and enums (not part of __all__, but public in each module)
Navigation

Public classes

Class	Type	Backend	API key needed	Page
`BaseEmbedder`	Abstract base class	—	—	base_embedder
`EmbedderConfig`	Dataclass	—	—	config_embedder
`ArabicEmbedder`	Concrete	Local (`sentence-transformers`)	No	arabic_embedder
`HuggingFaceEmbedder`	Concrete	Local (`sentence-transformers`)	No	huggingface_embedder
`OpenAIEmbedder`	Concrete	Hosted API	Yes (`OPENAI_API_KEY`)	openai_embedder
`GeminiEmbedder`	Concrete	Hosted API	Yes (`GOOGLE_API_KEY`/`GEMINI_API_KEY`)	gemini_embedder
`MistralEmbedder`	Concrete	Hosted API	Yes (`MISTRAL_API_KEY`)	mistral_embedder
`OllamaEmbedder`	Concrete	Local HTTP server	No	ollama_embedder

Common interface (from `BaseEmbedder`)

Every concrete embedder implements at minimum:

Member	Kind	Notes
`encode(texts, ...)`	abstract method	Must be implemented by every subclass; signatures vary slightly per class
`embedding_dim`	abstract property	Must be implemented by every subclass
`encode_with_cache(texts, **kwargs)`	method	Generic MD5-keyed cache wrapper around `encode()`
`similarity(text1, text2, metric)`	method	`'cosine'`, `'dot'`, or `'euclidean'`
`batch_similarity(query, texts, top_k)`	method	Dot-product similarity of a query against many texts
`validate_connection(test_text, detailed)`	method	Sanity-checks that `encode()` works; overridden by `OllamaEmbedder`
`get_model_info()`	method	Overridden by every subclass with backend-specific fields
`get_stats()` / `reset_stats()`	methods	Encoding count/time/cache-hit bookkeeping
`clear_cache()`	method	Overridden by most subclasses for backend-specific cache structures
`cleanup()`	method	Clears cache, frees GPU memory if `torch` + CUDA present
`timing(operation)`	context manager	Logs elapsed wall time for a `with` block
`aencode`, `aencode_with_cache`, `abatch_similarity`	async methods	Thread-pool wrappers via `asyncio.to_thread`
`__enter__`/`__exit__`, `__aenter__`/`__aexit__`	context managers	Sync and async; every subclass overrides the sync pair to log a summary
`__repr__`	dunder	Overridden by every subclass

Data classes and enums (not part of `all`, but public in each module)

Name	File	Purpose
`ArabicProcessingStats`	`arabic_embedder.py`	Tracks normalization counts/timing for `ArabicEmbedder`
`GeminiUsageStats`	`gemini_embedder.py`	Tracks request/character counts for `GeminiEmbedder`
`MistralUsageStats`	`mistral_embedder.py`	Tracks token/cost counts for `MistralEmbedder`
`MistralRateLimiter`	`mistral_embedder.py`	Simple sliding-window requests/second limiter
`UsageStats`	`openai_embedder.py`	Tracks token/cost/per-minute counts for `OpenAIEmbedder`
`RateLimiter`	`openai_embedder.py`	Sliding-window requests/tokens-per-minute limiter
`ArabicQuality`	`hugginface_embedder.py`	`Enum`: `EXCELLENT`, `GOOD`, `FAIR`, `UNKNOWN`
`ModelInfo`	`hugginface_embedder.py`	Dataclass describing a recommended HF model
`EmbeddingStats`	`hugginface_embedder.py`	Dataclass replacing `BaseEmbedder`'s dict `_stats` for `HuggingFaceEmbedder`
`retry_on_failure`	`hugginface_embedder.py`	Decorator factory implementing retry-with-backoff; used on `_load_model`

These aren't re-exported by embeddings/__init__.py, so import them from their submodule directly if needed, e.g. from embeddings.hugginface_embedder import ArabicQuality.

Related pages