constella-zero

constella-zero is a small English query encoder for semantic search. It uses an int8 token lookup table instead of a transformer, making it useful when query latency matters more than maximum retrieval quality.

It produces normalized 1024-dimensional vectors that search documents encoded by stella-en-400M-v5-doc-onnx. The same document index also works with the stronger constella-nano query encoder.

Research preview: Native FastEmbed support currently requires the Constella preview branch shown below. The published evaluation is limited to the results described in this card. Full BEIR benchmarking is underway. Results from that broader evaluation are not included yet.

The Constella family

The name Constella combines "constellation" and "Stella." The document embeddings are the fixed stars, and the query encoder navigates their shared vector space.

Zero and Nano are swappable at query time. Both can search the same document index, so you can choose between them without re-encoding documents or rebuilding the collection. Their rankings differ: Zero is the faster option, while Nano has higher retrieval scores on the six reported datasets. The "zero" name refers to its transformer-free query path.

Installation

Native FastEmbed support is currently available from the Constella preview branch:

pip install "fastembed>=0.9.0"
pip install qdrant-client

Usage

Encode documents once with the document model, then encode queries with constella-zero. The example below creates an in-memory Qdrant collection, but the vectors can be used with any vector database that supports cosine similarity.

from fastembed import TextEmbedding
from qdrant_client import QdrantClient, models

NAME = "Qdrant/constella-zero"
DOC_NAME = "Qdrant/stella-en-400M-v5-doc-onnx"

documents = [
    "mRNA vaccines deliver messenger RNA encoding a viral antigen.",
    "The Treaty of Westphalia ended the Thirty Years' War in 1648.",
]

document_model = TextEmbedding(DOC_NAME)

client = QdrantClient(":memory:")
client.create_collection(
    "documents",
    vectors_config=models.VectorParams(size=1024, distance=models.Distance.COSINE),
)
client.upsert(
    "documents",
    points=[
        models.PointStruct(id=i, vector=embedding.tolist(), payload={"text": text})
        for i, (text, embedding) in enumerate(
            zip(documents, document_model.embed(documents))
        )
    ],
)

query_model = TextEmbedding(NAME)
query_embedding = next(iter(query_model.embed(["how do mRNA vaccines work?"])))
results = client.query_points(
    "documents", query=query_embedding.tolist(), limit=2
).points

for result in results:
    print(result.score, result.payload["text"])

FastEmbed handles pooling and L2 normalization. Do not use the document model as an unprompted query encoder. Use constella-zero, constella-nano, or Stella's prompted query path instead.

NumPy reference implementation

The repository also includes zero_encoder.py, a reference implementation that does not require FastEmbed or ONNX Runtime:

from huggingface_hub import snapshot_download
import sys

model_directory = snapshot_download("Qdrant/constella-zero")
sys.path.insert(0, model_directory)

from zero_encoder import ZeroQueryEncoder

model = ZeroQueryEncoder(model_directory, variant="int8")
query_embeddings = model.encode(["how do mRNA vaccines work?"])

How it works

The encoder tokenizes each query with WordPiece, looks up a learned vector for every token, and combines those vectors into one query embedding. Repeated tokens receive diminishing weight: a token that occurs c times contributes a total weight of sqrt(c). The result is L2-normalized.

This is a bag-of-tokens model. It does not represent word order, syntax, or negation directly. Learned token weights are already included in the table.

Retrieval results

The table reports exact-search nDCG@10. ArguAna and FiQA are marked because the Stella teacher discloses training or evaluation contact with those datasets. Results on those two datasets should therefore be interpreted separately from the other four.

Note: Stella discloses training or evaluation contact with ArguAna and FiQA.

The recommended deployment setup for Zero is hybrid retrieval. Retrieve with both Zero and BM25, then combine their results with Qdrant's distribution-based score fusion (DBSF), prefetching 100 candidates from each side. This setup scored 0.4887 mean nDCG@10 across all six datasets and 0.4912 across the four datasets without disclosed Stella contact. The evaluated lexical side used bm25s with Lucene defaults, so results may differ with another BM25 implementation. Dense-only retrieval remains supported when a lexical index is unavailable or unnecessary.

Query encoding cost

These measurements cover the query encoder only. They use batch size 1, four CPU threads, five warmups, and twenty synthetic 20-word queries in each of three fresh processes. They do not include vector search or end-to-end application latency.

Files

Both ONNX graphs use opset 17 and standard operators. The int8 table is dequantized inside the graph with one fp32 scale per row.

Training

The released table was trained by L2 regression against Stella query embeddings. Training used 338,076 usable query-document pairs plus 220,632 query-only rows from Amazon ESCI, FEVER, HotpotQA, SQuAD, NQ Open, TriviaQA, and Mr. TyDi English. MS MARCO was excluded.

Wikipedia-derived data retains CC BY-SA attribution. Amazon ESCI and TriviaQA are Apache-2.0.

Limitations

• The model is English-only and truncates inputs after 512 tokens.
• As a bag-of-tokens model, it is weak at distinctions that depend on word order, syntax, or negation.
• Document indexing still requires the 400M-parameter Stella document encoder.

License and provenance

The model is MIT licensed. It was distilled from NovaSearch/stella_en_400M_v5 at revision ffeb2b7ee715c226d4ffe5e4619f7dbb48624c20, which is also MIT licensed.

The released int8 table has SHA-256 a7007b1a6af120b976f093fd69ddcb5001996ec0b84b5864b4fd25d7af878abf.