Embedding Models

Embedding Models

Decision Framework

Is text English-only and budget sensitive?
  YES → text-embedding-3-small (1536) or nomic-embed-text-v1.5
  NO  → ↓

Multilingual required?
  YES → BGE-M3, multilingual-e5-large, or cohere embed-multilingual-v3
  NO  → ↓

Retrieval quality top priority (leaderboard chasing)?
  YES → voyage-3-large, cohere embed-v3, or BGE-reranker + BGE-base
  NO  → text-embedding-3-large (3072, truncatable via MRL)

Self-hosting required (data residency)?
  YES → BGE-M3, E5-large-v2, nomic-embed-text-v1.5, mxbai-embed-large-v1

Model Catalog (2025)

MTEB numbers drift — verify at https://huggingface.co/spaces/mteb/leaderboard before committing to a model.

OpenAI

from openai import OpenAI

client = OpenAI()

def embed_openai(texts: list[str], model: str = "text-embedding-3-small",
                 dimensions: int | None = None) -> list[list[float]]:
    kwargs = {"model": model, "input": texts}
    if dimensions is not None:  # MRL truncation, only for -3 family
        kwargs["dimensions"] = dimensions
    resp = client.embeddings.create(**kwargs)
    return [d.embedding for d in resp.data]

# Cheap path: 3-small at default 1536
doc_vecs = embed_openai(docs)

# Storage-optimised: 3-large truncated to 512
compact = embed_openai(docs, model="text-embedding-3-large", dimensions=512)

Voyage AI

import voyageai

vo = voyageai.Client()

# Note the input_type — Voyage optimises query vs document embeddings separately
doc_vecs = vo.embed(docs, model="voyage-3", input_type="document").embeddings
qry_vec  = vo.embed([query], model="voyage-3", input_type="query").embeddings[0]

# Code embeddings
code_vecs = vo.embed(snippets, model="voyage-code-3", input_type="document").embeddings

Cohere

import cohere

co = cohere.ClientV2()

doc_resp = co.embed(
    texts=docs,
    model="embed-english-v3.0",
    input_type="search_document",
    embedding_types=["float"],
)
qry_resp = co.embed(
    texts=[query],
    model="embed-english-v3.0",
    input_type="search_query",
    embedding_types=["float"],
)
doc_vecs = doc_resp.embeddings.float
qry_vec  = qry_resp.embeddings.float[0]

# Binary embeddings for 32x storage reduction (see vector-quantization skill)
bin_resp = co.embed(
    texts=docs, model="embed-english-v3.0",
    input_type="search_document", embedding_types=["binary"],
)

BGE-M3 (self-hosted, OSS)

from FlagEmbedding import BGEM3FlagModel

model = BGEM3FlagModel("BAAI/bge-m3", use_fp16=True)

out = model.encode(
    docs,
    batch_size=12,
    max_length=8192,
    return_dense=True,
    return_sparse=True,       # lexical weights
    return_colbert_vecs=True, # multi-vector
)
dense   = out["dense_vecs"]         # (N, 1024)
sparse  = out["lexical_weights"]    # token-id -> weight
colbert = out["colbert_vecs"]       # per-token contextual vectors

E5 family (instruction-tuned)

E5 models require task prefixes — forget them and quality drops hard.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("intfloat/multilingual-e5-large")

# E5 REQUIRES these prefixes
doc_vecs = model.encode([f"passage: {d}" for d in docs], normalize_embeddings=True)
qry_vec  = model.encode([f"query: {query}"], normalize_embeddings=True)[0]

Jina v3 (task-specific LoRA)

from transformers import AutoModel

model = AutoModel.from_pretrained("jinaai/jina-embeddings-v3", trust_remote_code=True)

doc_vecs = model.encode(docs, task="retrieval.passage")
qry_vec  = model.encode([query], task="retrieval.query")[0]
# Other tasks: separation, classification, text-matching

Nomic (open weights, fully reproducible)

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("nomic-ai/nomic-embed-text-v1.5", trust_remote_code=True)
doc_vecs = model.encode([f"search_document: {d}" for d in docs])
qry_vec  = model.encode([f"search_query: {query}"])[0]

mixedbread mxbai

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("mixedbread-ai/mxbai-embed-large-v1")
doc_vecs = model.encode(docs, normalize_embeddings=True)
# Query prompt recommended
qry_vec = model.encode(
    [query],
    prompt="Represent this sentence for searching relevant passages: ",
    normalize_embeddings=True,
)[0]

Query vs Document Modes

Asymmetric models (Cohere, Voyage, E5, Nomic, BGE-M3 with prompts) encode queries and documents differently. Using the wrong mode costs 5-15% recall.

Cost Modelling

def monthly_cost(docs: int, avg_tokens: int, qps: int,
                 doc_rate_per_1m: float, qry_rate_per_1m: float) -> float:
    doc_tokens_once = docs * avg_tokens                     # one-time index
    qry_tokens_monthly = qps * 60 * 60 * 24 * 30 * 20       # ~20 tok/query
    return (doc_tokens_once * doc_rate_per_1m
            + qry_tokens_monthly * qry_rate_per_1m) / 1_000_000

# 1M docs x 500 tokens, 10 QPS on OpenAI 3-small ($0.02/1M)
print(monthly_cost(1_000_000, 500, 10, 0.02, 0.02))  # ~$10.4

Anti-Patterns

Production Checklist

• Model + version pinned in config and stored in vector DB metadata
• Query/document mode set correctly for asymmetric models
• Max token budget matches chunk size
• Private eval set scored before swapping models
• Cost model calculated for index + query volume
• Rate limits and batch sizes tuned (100-2048 per call)
• Re-index plan documented for model upgrades
• Fallback to secondary provider on API outage