Add API-accessible reranker feature using HuggingFaceCrossEncoder

[practicustalhasahin]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

I’d like to propose adding a feature to expose the HuggingFace-based reranker (HuggingFaceCrossEncoder) class via an API-accessible interface, similar to how embedding models or LLMs can be called through LangChain APIs.

Motivation

This would enable:

• External services to query reranker models via a standardized endpoint.
• Seamless integration of cross-encoder models (e.g., BAAI/bge-reranker-base) in reranking pipelines via REST or LangServe.
• More flexible use in retrieval + reranking pipelines, especially in production systems.

Proposal (If applicable)

Proposed change: Expose the scoring function of HuggingFaceCrossEncoder via a FastAPI or LangServe-compatible route.

Reference Class:

from langchain_community.cross_encoders import HuggingFaceCrossEncoder

hf = HuggingFaceCrossEncoder(
    model_name="BAAI/bge-reranker-base",
    model_kwargs={'device': 'cpu'}
)

scores = hf.score([("query", "doc1"), ("query", "doc2")])

Suggested additions:

• API wrapper class (e.g., BaseCrossEncoderAPI)
• Scoring route: /rerank/score
• Input format: {"text_pairs": [["query", "doc1"], ["query", "doc2"]]}

Use cases:

• Reranking results from retrieval.
• Enhancing search quality in RAG systems.
• Supporting hybrid search pipelines over LangChain-compatible APIs.