Models

SIE supports cross-encoder rerankers and ColBERT-style multivector models for reranking. Model performance varies by task. Run mise run eval <model> -t <task> to benchmark on your data.

---

Cross-Encoder Models

Cross-encoders score query-document pairs with full cross-attention. They see both inputs together for deeper semantic matching.

BGE Rerankers

Model	Max Length	Notes
BAAI/bge-reranker-base	512	Smaller, English-focused
BAAI/bge-reranker-large	512	Larger, English-focused
BAAI/bge-reranker-v2-m3	8192	Long context, 100+ languages

Mixedbread Rerankers

Model	Max Length	Notes
mixedbread-ai/mxbai-rerank-base-v2	8192	Base size
mixedbread-ai/mxbai-rerank-large-v2	8192	Larger model

Jina Rerankers

Model	Max Length	Notes
jinaai/jina-reranker-v2-base-multilingual	8192	100+ languages

GTE Rerankers

Model	Max Length	Notes
Alibaba-NLP/gte-reranker-modernbert-base	8192	ModernBERT architecture

MS-MARCO Cross-Encoders

Model	Max Length	Notes
cross-encoder/ms-marco-MiniLM-L-6-v2	512	Smallest, fastest
cross-encoder/ms-marco-MiniLM-L-12-v2	512	Balanced

---

Multi-Vector Rerankers (ColBERT)

ColBERT models produce per-token embeddings and score with MaxSim. Useful when you have pre-computed multivector embeddings.

Model	Token Dim	Max Length	Notes
colbert-ir/colbertv2.0	128	512	Original ColBERT
jinaai/jina-colbert-v2	128	8192	Long context, multilingual
mixedbread-ai/mxbai-colbert-large-v1	128	512	Larger model
mixedbread-ai/mxbai-edge-colbert-v0-32m	64	8192	Compact, edge-friendly
answerdotai/answerai-colbert-small-v1	96	512	Compact
lightonai/GTE-ModernColBERT-v1	128	8192	ModernBERT architecture
lightonai/Reason-ModernColBERT	128	8192	Reasoning-focused
nvidia/llama-nemoretriever-colembed-3b-v1	128	8192	Text + image, LLM-based

MaxSim Scoring

ColBERT reranking uses MaxSim over pre-encoded multivector embeddings:

Python

from sie_sdk import SIEClient
from sie_sdk.types import Item
from sie_sdk.scoring import maxsim

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

# Encode query and documents
query_result = client.encode(
    "jinaai/jina-colbert-v2",
    Item(text="What is ColBERT?"),
    output_types=["multivector"],
    is_query=True,
)

doc_results = client.encode(
    "jinaai/jina-colbert-v2",
    [Item(text="ColBERT uses late interaction..."), Item(text="The weather is nice.")],
    output_types=["multivector"]
)

# Score with MaxSim
query_mv = query_result["multivector"]
doc_mvs = [r["multivector"] for r in doc_results]
scores = maxsim(query_mv, doc_mvs)

# Rank by score
ranked = sorted(enumerate(scores), key=lambda x: -x[1])

TypeScript

import { SIEClient, maxsim } from "@sie/sdk";

const client = new SIEClient("http://localhost:8080");

// Encode query and documents
const queryResult = await client.encode(
  "jinaai/jina-colbert-v2",
  { text: "What is ColBERT?" },
  { outputTypes: ["multivector"], isQuery: true }
);

const docResults = await client.encode(
  "jinaai/jina-colbert-v2",
  [
    { text: "ColBERT uses late interaction..." },
    { text: "The weather is nice." },
  ],
  { outputTypes: ["multivector"] }
);

// Score with MaxSim using SDK helper
const queryMv = queryResult.multivector!;
const scores = docResults.map((r) => maxsim(queryMv, r.multivector!));

// Rank by score
const ranked = scores
  .map((score, idx) => ({ idx, score }))
  .sort((a, b) => b.score - a.score);

await client.close();

---

Model Selection

By Language Support

English only:

• BAAI/bge-reranker-base, BAAI/bge-reranker-large
• cross-encoder/ms-marco-MiniLM-L-6-v2, cross-encoder/ms-marco-MiniLM-L-12-v2

Multilingual (100+ languages):

• BAAI/bge-reranker-v2-m3
• jinaai/jina-reranker-v2-base-multilingual
• jinaai/jina-colbert-v2

By Context Length

Short context (512 tokens):

• BAAI/bge-reranker-base, BAAI/bge-reranker-large
• cross-encoder/ms-marco-MiniLM-L-*
• colbert-ir/colbertv2.0, mixedbread-ai/mxbai-colbert-large-v1

Long context (8192 tokens):

• BAAI/bge-reranker-v2-m3
• jinaai/jina-reranker-v2-base-multilingual
• Alibaba-NLP/gte-reranker-modernbert-base
• mixedbread-ai/mxbai-rerank-base-v2, mixedbread-ai/mxbai-rerank-large-v2
• jinaai/jina-colbert-v2, lightonai/GTE-ModernColBERT-v1

By Size

Compact (fast inference):

• cross-encoder/ms-marco-MiniLM-L-6-v2 - smallest cross-encoder
• mixedbread-ai/mxbai-edge-colbert-v0-32m - 32M parameters
• answerdotai/answerai-colbert-small-v1 - compact ColBERT

Large (higher capacity):

• BAAI/bge-reranker-large
• mixedbread-ai/mxbai-rerank-large-v2
• mixedbread-ai/mxbai-colbert-large-v1
• nvidia/llama-nemoretriever-colembed-3b-v1 - 3B parameters

---

Benchmarking

Use the eval harness to benchmark rerankers on your data:

# Quality evaluation
mise run eval BAAI/bge-reranker-v2-m3 -t mteb/AskUbuntuDupQuestions --type quality

# Performance evaluation
mise run eval BAAI/bge-reranker-v2-m3 -t mteb/AskUbuntuDupQuestions --type perf

# Compare multiple models
mise run eval BAAI/bge-reranker-base -t mteb/AskUbuntuDupQuestions --type quality
mise run eval cross-encoder/ms-marco-MiniLM-L-12-v2 -t mteb/AskUbuntuDupQuestions --type quality

---

What’s Next

• Rerank guide - how to use reranking in your pipeline
• Multi-vector embeddings - ColBERT encoding details