Backends¶
Optional ML / LLM backends implementing the pluggable detector interfaces.
Each backend lives in a submodule and requires its own optional extra
([llm], etc.). Core installs are unaffected — backends import lazily.
Embeddings similarity¶
EmbeddingsSimilarity(*, model_name: str | None = None, device: str | None = None, cache_size: int = 10000, model: _Encoder | None = None)
Cosine-similarity backend using sentence-transformer embeddings.
This class is callable — instances conform to the SimilarityFn protocol
and can be passed directly to DeathLoopDetector(similarity_fn=...).
Embeddings are cached by text so repeated comparisons against the same response don't re-run the model.
Initialize the backend.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
model_name
|
str | None
|
Any Hugging Face sentence-transformers model id.
Defaults to |
None
|
device
|
str | None
|
|
None
|
cache_size
|
int
|
Maximum distinct texts to keep embeddings for. LRU. |
10000
|
model
|
_Encoder | None
|
Pre-built encoder instance. If provided, |
None
|