stella-en-400M-v5-doc-onnx

stella-en-400M-v5-doc-onnx

This is the document encoder used by the Constella asymmetric retrieval models. It converts English documents into normalized 1024-dimensional vectors. Index documents with this model, then search the index with either constella-zero or constella-nano.

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

The model is an ONNX conversion of NovaSearch/stella_en_400M_v5 at revision ffeb2b7ee715c226d4ffe5e4619f7dbb48624c20. The weights were not trained or fine-tuned during conversion. Pooling and normalization are included in the graph, and inference does not require PyTorch or trust_remote_code.

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 documents encoded by this model, so you can change query encoders without re-encoding documents or rebuilding the collection. Compatibility refers to the shared vector space; rankings and retrieval quality vary by query encoder.

Installation

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

pip install "fastembed>=0.9.0"

Usage

from fastembed import TextEmbedding

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

documents = [
    "Marie Curie conducted pioneering research on radioactivity.",
    "The Nile is a major north-flowing river in northeastern Africa.",
]

model = TextEmbedding(NAME)
embeddings = list(model.embed(documents))

Store the returned vectors in a cosine-similarity index. Use constella-zero or constella-nano to encode queries against that index.

This graph is the document path only. If you use the original Stella model for queries, use its s2p_query prompt as described on the source model card. Do not use this artifact as an unprompted Stella query encoder.

How it works

The graph applies Stella's document path and then computes:

masked mean over last_hidden_state -> 2_Dense_1024 -> L2 normalization

Inputs are int64 input_ids and attention_mask. The output is a normalized fp32 array with shape (batch, 1024). Inputs are truncated at 512 tokens. Paired sequences and token_type_ids are not supported.

Conversion accuracy

The ONNX graph was compared with the original PyTorch document path on 259 Natural Questions passages. The fixtures cover short, medium, boundary-length, and over-length inputs.

Output norms were 0.999999 to 1.000001. Encoding the same text alone or in a ragged batch produced bit-identical results.

Files

The graph uses ONNX opset 17, standard operators, and no external-data initializers. The tokenizer uses dynamic padding and truncates at 512 tokens.

An fp16 graph is not included. The tested fp16 conversion reached a minimum cosine similarity of 0.662 against the fp32 reference on CUDA and was not accurate enough to release.

Training

No training, fine-tuning, or distillation was performed for this conversion. The graph contains the source Stella weights and the same document-side computation in ONNX format.

Limitations

• This artifact encodes documents only. It does not apply Stella's query prompt.
• The model is English-only and truncates inputs after 512 tokens.
• The 400M-parameter encoder and 1.75 GB graph are intended for indexing, not lightweight query serving.
• Compatibility with the Constella query encoders means they share a vector space. It does not mean the query encoders have equal retrieval quality.

License and provenance

NovaSearch releases the pinned Stella weights under the MIT license. This repository redistributes the same weights in ONNX form under that license. Stella derives from Alibaba-NLP/gte-large-en-v1.5, which is Apache-2.0 licensed.

The graph was exported with torch.onnx.export, opset 17, and constant folding. Its SHA-256 is fe31555e2b40767e17487885fb67dcdf0dcee11bef31f42478e55c1ec69a4ea9.