Search long documents 79% cheaper than voyage‑context‑4

GTE ModernColBERT turns whole manuals, contracts and reports into token vectors, so individual details can match a question. SIE is the open-source inference engine that serves the model through one API.

Get started Run in your cloud

SIE GTE ModernColBERT

Right article first of 11,802

Question

What rank was Sheehan granted when he was decommissioned?

D. D. Sheehan 8,190 tokens, indexed whole

Matched paragraph, 59% of the way through

… Deafness by shellfire and ill-health necessitated his transfer to the 3rd RMF (Reserve) Battalion at Aghada, then Ballincollig barracks, Co. Cork. Hospitalised often, he was decommissioned late 1917, with a bulletin stating that he "relinquished his commission on account of ill-health contracted on active service, and is granted the permanent honorary rank of Captain, 13 Jan.1918".The Cork Constitution newspaper notice: 15 January 1918 at Cork City Council Library Sheehan was awarded the World War I campaign medals: 1914–15 Star, British War Medal, Victory Medal and Silver War Badge. …

Wikipedia, D. D. Sheehan · CC BY-SA Marked: where the question’s tokens matched best

The answer can fall outside the retrieved context

OpenAI text-embedding-3-large, chunked

Wrong article first

Question

What rank was Sheehan granted when he was decommissioned?

Ranked 1st: Carroll Sheehan Wikipedia · CC BY-SA

Carroll P. Sheehan (1925-2016) was an American real estate executive and political figure who was a candidate for Governor of Massachusetts in 1974.

The right article, D. D. Sheehan, ranked 2nd. A reader opening the first result gets the wrong person.

voyage-context-4

Answer article ranks fifth

Question

According to Harriet Hanson Robinson, what was the purpose of offering high wages to the Lowell mill girls?

Ranked 5th: Wage slavery Wikipedia · CC BY-SA

In an account of the Lowell mill girls, Harriet Hanson Robinson wrote that generously high wages were offered to overcome the degrading nature of the work:

The labelled answer article ranks fifth. Retrieving only the first article leaves this passage out.

Voyage AI A clause embedded alone loses the definitions it depends on
Standard embedding models encode each chunk independently, discarding the surrounding document context
29 June 2026
Anthropic Retrieval can fail when chunks lose their surrounding context
traditional RAG solutions remove context when encoding information
19 September 2024

96% of voyage‑context‑4’s accuracy, 79% cheaper

Run this eval

Right article first

of 437 questions ↑

voyage-context-4

voyage-context-4 280

SIE GTE ModernColBERT

SIE GTE ModernColBERT 269

Voyage 4 Large

Voyage 4 Large 243

OpenAI text-embedding-3-large

OpenAI text-embedding-3-large 228

Cohere Embed v5 Pro

Cohere Embed v5 Pro 193

Index 11,802 articles

USD ↓

SIE GTE ModernColBERT

$2.11 SIE GTE ModernColBERT

Cohere Embed v5 Pro

Cohere Embed v5 Pro $9.47

Voyage 4 Large

Voyage 4 Large $9.88

voyage-context-4

voyage-context-4 $10.15

OpenAI text-embedding-3-large

OpenAI text-embedding-3-large $12.76

Evaluation

Tested: 437 held-out MLDR questions, 11,802 English Wikipedia articles
Scored: The labelled answer article ranks first
Prices: SIE Serverless and provider list prices; each model’s billed tokens

Send a question and its passages, get token vectors to rank

View on GitHub

from sie_sdk import SIEClient
client = SIEClient(api_key="API keysk-sie-…")
query = "querycookie sent with cross-site requests that also needs the Secure attribute"
passages = [
    "passage 1Send the cookie only for requests originating from the same site that set the cookie.",
    "passage 2Send the cookie only for requests originating from the same site that set the cookie, and for cross-site requests that meet both of the following criteria:",
    "passage 3Send the cookie with both cross-site and same-site requests. The Secure attribute must also be set when using this value.",
    "passage 4Indicates that the cookie should be stored using partitioned storage. Note that if this is set, the Secure directive must also be set.",
]
q = client.encode(
    "modellightonai/GTE-ModernColBERT-v1",
    [{"text": query}],
    output_types=["multivector"],
    is_query=True,
)[0]["multivector"]
docs = client.encode(
    "modellightonai/GTE-ModernColBERT-v1",
    [{"text": passage} for passage in passages],
    output_types=["multivector"],
    is_query=False,
)
print(q)
print(docs)

import { SIEClient } from '@superlinked/sie-sdk';

const client = new SIEClient(undefined, { apiKey: 'sk-sie-…' });

const query = "cookie sent with cross-site requests that also needs the Secure attribute";
const passages = [
  "Send the cookie only for requests originating from the same site that set the cookie.",
  "Send the cookie only for requests originating from the same site that set the cookie, and for cross-site requests that meet both of the following criteria:",
  "Send the cookie with both cross-site and same-site requests. The Secure attribute must also be set when using this value.",
  "Indicates that the cookie should be stored using partitioned storage. Note that if this is set, the Secure directive must also be set.",
];
const model = 'lightonai/GTE-ModernColBERT-v1';

const [q] = await client.encode(model, [{ text: query }], {
  outputTypes: ['multivector'], isQuery: true,
});
const docs = await client.encode(model, passages.map((text) => ({ text })), {
  outputTypes: ['multivector'], isQuery: false,
});
console.log({ query: q, documents: docs });

# Encode the question as a query, then the passages as documents; rank by MaxSim.
curl https://api.superlinked.com/v1/encode/lightonai%2FGTE-ModernColBERT-v1 \
  -H "Authorization: Bearer sk-sie-…" \
  -H "Content-Type: application/json" \
  -d "{\"items\":[{\"text\":\"cookie sent with cross-site requests that also needs the Secure attribute\"}],\"params\":{\"output_types\":[\"multivector\"],\"options\":{\"is_query\":true}}}"

curl https://api.superlinked.com/v1/encode/lightonai%2FGTE-ModernColBERT-v1 \
  -H "Authorization: Bearer sk-sie-…" \
  -H "Content-Type: application/json" \
  -d "{\"items\":[{\"text\":\"Send the cookie only for requests originating from the same site that set the cookie.\"},{\"text\":\"Send the cookie only for requests originating from the same site that set the cookie, and for cross-site requests that meet both of the following criteria:\"},{\"text\":\"Send the cookie with both cross-site and same-site requests. The Secure attribute must also be set when using this value.\"},{\"text\":\"Indicates that the cookie should be stored using partitioned storage. Note that if this is set, the Secure directive must also be set.\"}],\"params\":{\"output_types\":[\"multivector\"],\"options\":{\"is_query\":false}}}"

Build the "Multi-vector" capability into my app using the Superlinked Inference Engine (SIE).

Context
- SIE is an OpenAI-style inference API. Python SDK: `from sie_sdk import SIEClient`; TypeScript: `@superlinked/sie-sdk`.
- Base URL: https://api.superlinked.com (or my regional endpoint). Auth: Bearer key from env `SIE_API_KEY` (never hard-code it).
- Model: lightonai/GTE-ModernColBERT-v1 (SIE primitive: /encode). Keep the model id configurable.

Task
- Input: a query string plus a list of candidate documents.
- Behaviour: return the query and passage token vectors from the native encode calls
- Send two POST requests to https://api.superlinked.com/v1/encode/lightonai%2FGTE-ModernColBERT-v1. Encode one query item {text: query} with params {output_types: ["multivector"], options: {is_query: true}}, then all candidate passages in one batch of {text: passage} items with params {output_types: ["multivector"], options: {is_query: false}}. Do not call the API per candidate. Read the returned per-token vectors from items[i].multivector.values. Return the native token vectors without adding application scoring or ranking. In Python, use SIEClient.encode(..., output_types=["multivector"], is_query=True) for the query and encode(..., output_types=["multivector"], is_query=False) for the passage batch; read each decoded ["multivector"] array. In TypeScript, use client.encode with outputTypes: ["multivector"] and isQuery: true for the query, then isQuery: false for the passage batch; the SDK decodes the native token vectors.

Deliverables
- A typed client wrapper, an application-level function for this task, error handling for timeouts/empty input, and unit tests with a stubbed client.
- Wire it into my existing stack (ask me which framework if unclear) and add a short usage example.

Run

Output

RankPassageMaxSim

SameSite=None

Send the cookie with both cross-site and same-site requests. The Secure attribute must also be set when using this value.

25 token vectors × 128

15.089

SameSite=Lax

Send the cookie only for requests originating from the same site that set the cookie, and for cross-site requests that meet both of the following criteria:

30 token vectors × 128

14.742

SameSite=Strict

Send the cookie only for requests originating from the same site that set the cookie.

18 token vectors × 128

14.609

Partitioned

Indicates that the cookie should be stored using partitioned storage. Note that if this is set, the Secure directive must also be set.

28 token vectors × 128

14.594

The API returns 16 query token vectors and each passage’s token vectors. This display ranks the recorded vectors with MaxSim. Recorded 2026-09-15; MDN Web Docs, CC BY-SA 2.5.

Index long documents 79% cheaper than voyage‑context‑4

Serverless API

We host the models; you call one API.

No idle GPUs, pay for what you use
Fits your stack: SDK, API, CLI, MCP
Self-host the same open-source engine
SOC 2 Type 2, US or EU data residency
Create an account to try Serverless

Get started

Self-hosted (Kubernetes)

Your data stays in your cloud, backed by our engineers.

Terraform modules for your cloud
Apache-2.0, same engine as Serverless
GPU workers scale to zero between jobs
Separate worker pools for each tenant

Agent prompt Deploy SIE to our AWS account with the superlinked/sie/aws Terraform module. Docs: superlinked.com/docs/deployment.

Deploy SIE to our Google Cloud project with the superlinked/sie/google Terraform module. Docs: superlinked.com/docs/deployment.

Deploy SIE to our Azure subscription with the superlinked/sie/azure Terraform module. Docs: superlinked.com/docs/deployment.

Deploy SIE to our Alibaba Cloud account with the superlinked/sie/alicloud Terraform module. Docs: superlinked.com/docs/deployment.

Talk to us Deploy guide

Run locally

The open-source server on your own machine.

NVIDIA
One container, no cluster
Process your data fully offline
Same SDK and model IDs as Serverless

docker run --gpus all -p 8080:8080 ghcr.io/superlinked/sie-server:latest-cuda12-default

Quickstart

Switch to SIE in 5 minutes

from openai import OpenAI

client = OpenAI()
model = "text-embedding-3-large"

docs = client.embeddings.create(model=model, input=document_chunks)
query = client.embeddings.create(model=model, input=[question])

from pylate import models

model = models.ColBERT("lightonai/GTE-ModernColBERT-v1", document_length=8192)

docs = model.encode(documents, is_query=False)
query = model.encode([question], is_query=True)

SIE

from sie_sdk import SIEClient

client = SIEClient(api_key="sk-sie-…")
model = "lightonai/GTE-ModernColBERT-v1"
options = {"profile": "long_context"}  # whole documents, 8,192-token limit

docs = client.encode(
    model, [{"text": d} for d in documents],
    output_types=["multivector"], is_query=False, options=options,
)
query = client.encode(
    model, [{"text": question}],
    output_types=["multivector"], is_query=True, options=options,
)

from sie_sdk import SIEClient

client = SIEClient(api_key="sk-sie-…")
model = "lightonai/GTE-ModernColBERT-v1"
options = {"profile": "long_context"}  # whole documents, 8,192-token limit

docs = client.encode(
    model, [{"text": d} for d in documents],
    output_types=["multivector"], is_query=False, options=options,
)
query = client.encode(
    model, [{"text": question}],
    output_types=["multivector"], is_query=True, options=options,
)