itsukifujii/kalm-reranker-v1-small-r2

Newest reranker model from KaLM-Embedding (https://huggingface.co/KaLM-Embedding/KaLM-Reranker-V1-Small-R2).

Public
5 runs
License
GitHub
Weights
Paper
Playground API Examples README Versions

Run time and cost

This model runs on Nvidia L40S GPU hardware. We don't yet have enough runs of this model to provide performance information.

Readme

KaLM-Reranker-V1-Small-R2

An open-source multilingual document reranker from the KaLM-Embedding Research Team, designed for search and retrieval-augmented generation (RAG).

KaLM uses an encoder–decoder architecture with compressed document representations. This endpoint exposes compression_ratio as a per-request hyperparameter, enabling users to explore test-time compute and ranking-quality trade-offs without retraining.

Adjustable test-time compute

Set compression_ratio when making a prediction:

  • 1: No document representation compression.
  • 4: Default compression.
  • 32, 64, or 128: More aggressive compression.

Higher compression reduces the number of document representations passed to the scoring decoder. You can change the compression ratio between requests while using the same model checkpoint.

Inputs

Parameter Description
query Query text. Required.
documents Array of candidate document texts. Required.
compression_ratio Integer from 1 to 128. Default: 4.
top_n Maximum number of results. Omit to return all documents.
return_documents Include document text in the output. Default: false.
instruction Optional task instruction to guide relevance scoring.

Example

{
  "query": "What is the capital of France?",
  "documents": [
    "Berlin is the capital of Germany.",
    "Paris is the capital of France.",
    "Rome is the capital of Italy.",
    "Madrid is the capital of Spain."
  ],
  "compression_ratio": 4,
  "top_n": 4,
  "return_documents": true
}

Try the same input with compression_ratio set to 32 to compare configurations.

Output

Results are sorted by descending relevance score. Each result contains:

  • index: Original document index, starting from 0.
  • relevance_score: Model relevance score.
  • document: Original text, when return_documents is true.
Model created