hs-rerank-1

hs-rerank-1

Scores how well each document answers a query, in eight languages and for inputs up to 8,192 tokens, so a search or memory system can put the best matches first. For developers building retrieval.

Made by hindsight-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: hs-rerank-1 is a small multilingual reranker built for the Hindsight project. It reads a query and a candidate document together and gives back one relevance score between 0 and 1. A search or memory system can use those scores to put the best matches at the top.

Who it suits: It suits developers who run search, retrieval or memory features and want a light model that sorts results across several languages.

Industry | Who uses it | How they use it | Example
Software | Search and retrieval engineers | Re-sort the top results from a first-stage search | Ranking candidate passages for a question
AI agents | Agent developers | Pick the stored memories that fit the current request | Choosing which notes to show a user
Support | Help desk tool builders | Match a customer question to the best article | Ranking help pages for a Spanish query
Research | Retrieval researchers | Use a compact baseline across languages | Comparing scores on mixed-language data

Skip it if you only need to embed text for first-stage search, since this model scores pairs and does not create embeddings.

Where to use it:

• Python code using the sentence-transformers library
• Hindsight itself, set to a local model or a Text Embeddings Inference server
• Text Embeddings Inference servers, with CPU and GPU images
• The hosted Hindsight Inference service, with its own API key

Use case | Who | What they get back
Reranking results | Search developers | Documents sorted by a 0 to 1 relevance score
Memory lookup | Agent builders | The stored facts that best match a question
Multilingual matching | Global product teams | Scores for queries and documents in different languages

How to set it up: You need these first.

• Python with pip
• The sentence-transformers library
• A CPU or a GPU, using the ONNX files on CPU and the half-precision weights on GPU

1. Install the sentence-transformers library with pip.
2. Optionally fetch the files ahead with the Hugging Face tools.
3. Load the model by name in a CrossEncoder.

pip install -U sentence-transformers

pip install -U huggingface_hub
hf download hindsight-ai/hs-rerank-1

How to use it: Create a CrossEncoder with the model name, then call rank with a query and your documents. Scores run from 0 to 1. Pass the identity activation to get raw logits instead. To use it inside Hindsight, set the environment variables below.

from sentence_transformers import CrossEncoder

model = CrossEncoder("hindsight-ai/hs-rerank-1")

query = "¿Qué app de presupuesto estoy creando?"
documents = [
    "El usuario quiere tooltips en su app de presupuesto.",
    "El usuario corre los domingos por la mañana.",
]

print(model.rank(query, documents))
# [{'corpus_id': 0, 'score': 0.9647}, {'corpus_id': 1, 'score': 0.0003}]

HINDSIGHT_API_RERANKER_PROVIDER=local
HINDSIGHT_API_RERANKER_LOCAL_MODEL=hindsight-ai/hs-rerank-1

Tool | What it does
rank | Scores each document against the query and sorts them
predict | Returns relevance scores between 0 and 1
activation_fn=torch.nn.Identity() | Returns raw logits in place of 0 to 1 scores
Text Embeddings Inference | Serves the model over HTTP on CPU or GPU

Example: Ask in Spanish which budget app you are building, and offer two notes. One says the user wants tooltips in their budget app, the other says they run on Sunday mornings. The first note scores 0.9647 and the second 0.0003.

For newcomers and experts: If you are new, a reranker is a second look at search results. A first search returns many candidates, and this model reads each one next to the question and scores it. Start with the short Python example above.

If you know the field, it is a cross-encoder based on mmBERT-small. Of its 114M parameters, 98M are the token embedding table, so each pass runs a 15M-parameter transformer. Inputs may reach 8,192 tokens per pair. The weights are stored in FP16, with an FP32 copy on a separate branch and an ONNX export with an int8 vocabulary table.

Query and one document go in together
Cross-encoder reads the pair
One relevance score comes out
Documents are sorted by score

Good to know:

• Languages: Chinese in both scripts, Japanese, Korean, English, and four European ones, Spanish, French, German and Portuguese
• Serving: Text Embeddings Inference has CPU and GPU images, and the GPU image tag depends on your card
• Hosted option: Hindsight Inference serves the model and needs its own API key

Title: Relevance scores in the card's Spanish example
Tooltips in a budget app: 0.9647 score
Runs on Sunday mornings: 0.0003 score

• License: MIT
• Cost to run: none named for local use, the hosted Hindsight Inference service needs a Hindsight Inference API key
• Data it reaches: the queries and documents you pass in, which stay local unless you choose a server or the hosted service
• Not for: creating embeddings, since it only scores a query and document pair

On Hugging Face, 13 November 2026

• Task: text ranking
• Library: sentence-transformers
• Parameters: 114M
• Weights: about 228 MB in F16; about 57 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,148,007
• Likes: 6
• License: mit
• Languages: en, es, fr, de, pt, ja, ko, zh
• Based on: jhu-clsp/mmBERT-small
• Tags: sentence-transformers, onnx, safetensors, modernbert, cross-encoder, reranker, multilingual, text-embeddings-inference, hindsight, text-ranking

Links

• Model: https://huggingface.co/hindsight-ai/hs-rerank-1
• Files: https://huggingface.co/hindsight-ai/hs-rerank-1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download hindsight-ai/hs-rerank-1

Then follow the usage in its model card. Nothing in this kit runs by itself.