agentforge-reranker-sentence-transformers

agentforge-reranker-sentence-transformers

SentenceTransformers cross-encoder reranker for the AgentForge framework (feat-021 default concrete impl).

Registers as agentforge.rerankers:sentence-transformers and implements the Reranker ABC from agentforge-core.

Installation

pip install agentforge-reranker-sentence-transformers[sentence-transformers]

The [sentence-transformers] extra pulls in sentence-transformers>=3.0 (~500MB for base + one model). Without it, the production factory raises ModuleError with pip remediation.

Usage

from agentforge import Retriever
from agentforge_reranker_sentence_transformers import (
    SentenceTransformersReranker,
)

reranker = SentenceTransformersReranker.from_config(
    model="cross-encoder/ms-marco-MiniLM-L-6-v2",
)
retriever = Retriever(
    store=vector_store,
    embedder=embedding_client,
    reranker=reranker,
    over_fetch_factor=3,
)
results = await retriever.retrieve("how do I deploy?", top_k=5)

Score normalisation

The cross-encoder returns raw logits (typically -10..+10). The reranker applies a sigmoid (1 / (1 + exp(-x))) so returned scores are normalised to (0, 1) — fitting the VectorMatch.score contract from agentforge-core.

License

Apache-2.0.