Cohere → SIE

Cohere offers embed-v3 (asymmetric query/document embeddings) and rerank-v3.5 (cross-encoder reranker). Both have direct SIE equivalents.

Why migrate

• Self-hosted. No API key, no per-call cost, no rate limits, no data egress.
• Embedding + reranking from one stack. Cohere bills both separately; SIE serves both from one cluster.
• Open weights. SIE serves permissive-licensed embedding models and BAAI/bge-reranker-v2-m3 from one self-hosted stack.

What stays the same

• Asymmetric query/document hint. Cohere uses input_type="search_query" vs "search_document". SIE uses is_query=True vs is_query=False.
• Rerank shape. Query + list of docs in, sorted scores out.
• Multilingual coverage. If you picked Cohere for embed-multilingual-v3.0, use a multilingual SIE model such as intfloat/multilingual-e5-large-instruct and validate on your own labeled multilingual set before cutting over; model rankings shift by language.

Before

import os
import cohere

docs = [
    "Mitochondria are the powerhouse of the cell.",
    "The Eiffel Tower is in Paris.",
]

client = cohere.ClientV2(api_key=os.environ["COHERE_API_KEY"])

resp = client.embed(
    texts=["The mitochondrion is the powerhouse of the cell."],
    model="embed-english-v3.0",
    input_type="search_document",
    embedding_types=["float"],
)
vectors = resp.embeddings.float_

resp = client.rerank(
    model="rerank-v3.5",
    query="What is the powerhouse of the cell?",
    documents=docs,
    top_n=10,
)

After

from sie_sdk import SIEClient
from sie_sdk.types import Item

client = SIEClient("http://localhost:8080")

texts = ["The mitochondrion is the powerhouse of the cell."]
docs = [
    "Mitochondria are the powerhouse of the cell.",
    "The Eiffel Tower is in Paris.",
]

# Document side: is_query=False (or omitted)
docs_result = client.encode(
    "NovaSearch/stella_en_400M_v5",
    [Item(text=t) for t in texts],
    is_query=False,
)
vectors = [r["dense"].tolist() for r in docs_result]

# Rerank
result = client.score(
    "BAAI/bge-reranker-v2-m3",
    Item(text="What is the powerhouse of the cell?"),
    [Item(text=d) for d in docs],
)
# result["scores"] is sorted by relevance desc; each entry has item_id and score.

Mapping

Cohere	SIE equivalent
embed-english-v3.0	NovaSearch/stella_en_400M_v5
embed-multilingual-v3.0	intfloat/multilingual-e5-large-instruct
embed-english-light-v3.0	intfloat/e5-small-v2
input_type="search_query"	client.encode(..., is_query=True)
input_type="search_document"	client.encode(..., is_query=False)
rerank-v3.5 / rerank-english-v3.0	BAAI/bge-reranker-v2-m3 or jinaai/jina-reranker-v2-base-multilingual
rerank-multilingual-v3.0	BAAI/bge-reranker-v2-m3

Score range caveat

Cohere returns calibrated relevance scores in roughly the [0, 1] range. SIE’s score() returns raw model-native scores from the cross-encoder, which for bge-reranker-v2-m3 are logits (any real number; sigmoid to get a probability if you need [0,1]). The ordering is what matters for reranking; don’t compare absolute scores across the two providers.

Re-embed required?

Yes. Cohere’s embed-v3 and SIE’s catalog embedding models produce vectors in unrelated spaces, even at matched dimensions.

Run it yourself

mise run serve -- -m NovaSearch/stella_en_400M_v5,BAAI/bge-reranker-v2-m3
export COHERE_API_KEY=...
uv add cohere

Run the Cohere ‘before’ and SIE ‘after’ snippets from this page on a small slice of your own corpus. Compare embeddings and rerank ordering.