Better page search than Voyage multimodal‑3.5, for 81% less

SIE runs open-source document search models behind one API. Search PDFs, slides and scanned reports straight from the page images, tables and charts included, with no OCR step to break.

Get started Run in your cloud

SIE Tomoro ColQwen3 4B

Right page

QUERY

What is the 2026 SEMIMONTHLY standard withholding for Married Filing Jointly with adjusted wages of $1,000?

1 · Page 20IRS Publication 15-T · 2026

SEMIMONTHLY Payroll Period

Table excerpt

Wage bracket $995 to $1,010; Married Filing Jointly standard withholding $0

Excerpt from the first of 8 ranked pages. Recorded query

OCR plus OpenAI embeddings returns the wrong page

OCR + OpenAI text-embedding-3-large

Wrong page first

Question

How much did the amount of JP Morgan Chase's total asset management fee increase from 2023 to 2024?

Open full source page ↗

Client assets, not fees. The answer page came 4th.

JPMorgan Chase 2024 Form 10-K

OCR + OpenAI text-embedding-3-large

Wrong page first

Question

What was Romania’s employment rate for people aged 15 to 24 in 2020?

Open full source page ↗

No figure for 2020 on this page. The answer page came 76th.

European Commission, Labour market and wage developments in Europe 2024 (CC BY 4.0)

arXiv:2412.02592 ICCV 2025 None of the OCR tools tested produced text good enough for a reliable RAG index
OCR Hinders RAG: Evaluating the Cascading Impact of OCR on Retrieval-Augmented Generation

Junyuan Zhang, Qintong Zhang, Bin Wang et al.

“none is competent for constructing high-quality knowledge bases for RAG systems”
3 December 2024

Better than Voyage multimodal‑3.5 for 81% less, level with Cohere Embed v5 Pro for 72% less

Run this eval

Ranking quality against price

1,816 questions

Page ranking quality (ViDoRe v3, of 100) ↑ Price per 1,000 pages →

ViDoRe v3 page ranking quality (nDCG@10, of 100) on 1,816 questions, against the real-time price per 1,000 pages indexed
Model	Page ranking quality (ViDoRe v3, of 100)	Price per 1,000 pages
SIE Tomoro ColQwen3 4B	64.1	$0.20
Cohere Embed v5 Pro	64.8	$0.74
Voyage multimodal-3.5	61.0	$1.11
OCR + OpenAI embeddings	56.0	$1.57

PDF page search on public documents

1,816 questions across 11,624 pages of public reports, filings and slides (ViDoRe v3)

Scored: Human relevance grades; nDCG@10, 100 is perfect
Prices: Real-time list prices, each at the page render its score came from

Encode the page once. Search it by text.

from pathlib import Path
from urllib.request import urlopen
from sie_sdk.scoring import maxsim
from sie_sdk import SIEClient
client = SIEClient(api_key="API keysk-sie-…")
model = "modelTomoroAI/tomoro-colqwen3-embed-4b:compact"
query = "queryWhat is the 2026 SEMIMONTHLY standard withholding for Married Filing Jointly with adjusted wages of $1,000?"
pages = [
    "irs-p15t-page-13.png", "irs-p15t-page-14.png",
    # Copy and Run include all 8 pages.
]
# Pinned image-loading setup omitted; Copy includes it.
q = client.encode(
    model, [{"text": query}],
    output_types=["multivector"], is_query=True,
)[0]["multivector"]
docs = client.encode(
    model, [{"images": [image]} for image in images],
    output_types=["multivector"], is_query=False,
)
# Native output is per-token vectors; your code ranks the pages.
scores = maxsim(q, [doc["multivector"] for doc in docs])
for i in sorted(range(len(pages)), key=lambda i: -scores[i]):
    print(f"{scores[i]:.3f}  {pages[i]}")

import { readFile } from 'node:fs/promises';
import { SIEClient } from '@superlinked/sie-sdk';

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

const model = 'TomoroAI/tomoro-colqwen3-embed-4b:compact';
const query = "What is the 2026 SEMIMONTHLY standard withholding for Married Filing Jointly with adjusted wages of $1,000?";
const pages = [
    "irs-p15t-page-13.png", "irs-p15t-page-14.png",
    // Copy and Run include all 8 pages.
];
// Pinned image-loading setup omitted; Copy includes it.

const [q] = await client.encode(
  model, [{ text: query }],
  { outputTypes: ['multivector'], isQuery: true },
);
const docs = await client.encode(
  model, images.map((image) => ({ images: [image] })),
  { outputTypes: ['multivector'], isQuery: false },
);
// Native output is per-token vectors; your code ranks the pages.
const dot = (a: Float32Array, b: Float32Array) => a.reduce((s, x, i) => s + x * b[i]!, 0);
const maxsim = (qv: Float32Array[], dv: Float32Array[]) =>
  qv.reduce((sum, qt) => sum + Math.max(...dv.map((dt) => dot(qt, dt))), 0);
const ranked = pages
  .map((page, i) => ({ page, score: maxsim(q!.multivector!, docs[i]!.multivector!) }))
  .sort((a, b) => b.score - a.score);
console.log(ranked);

page_0=$(curl --fail --silent --show-error --location 'https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/74f4e0eb2676a95b625783932796edac12040644/task-visual-examples-2026/irs-p15t-page-13.png' | base64 | tr -d '\n')
page_1=$(curl --fail --silent --show-error --location 'https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/74f4e0eb2676a95b625783932796edac12040644/task-visual-examples-2026/irs-p15t-page-14.png' | base64 | tr -d '\n')
page_2=$(curl --fail --silent --show-error --location 'https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/74f4e0eb2676a95b625783932796edac12040644/task-visual-examples-2026/irs-p15t-page-15.png' | base64 | tr -d '\n')
page_3=$(curl --fail --silent --show-error --location 'https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/74f4e0eb2676a95b625783932796edac12040644/task-visual-examples-2026/irs-p15t-page-16.png' | base64 | tr -d '\n')
page_4=$(curl --fail --silent --show-error --location 'https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/74f4e0eb2676a95b625783932796edac12040644/task-visual-examples-2026/irs-p15t-page-17.png' | base64 | tr -d '\n')
page_5=$(curl --fail --silent --show-error --location 'https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/74f4e0eb2676a95b625783932796edac12040644/task-visual-examples-2026/irs-p15t-page-18.png' | base64 | tr -d '\n')
page_6=$(curl --fail --silent --show-error --location 'https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/74f4e0eb2676a95b625783932796edac12040644/task-visual-examples-2026/irs-p15t-page-19.png' | base64 | tr -d '\n')
page_7=$(curl --fail --silent --show-error --location 'https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/74f4e0eb2676a95b625783932796edac12040644/task-visual-examples-2026/irs-p15t-page-20.png' | base64 | tr -d '\n')
curl https://api.superlinked.com/v1/encode/TomoroAI%2Ftomoro-colqwen3-embed-4b%3Acompact \
  -H "Authorization: Bearer sk-sie-…" \
  -H "Content-Type: application/json" \
  -d "{\"items\":[{\"text\":\"What is the 2026 SEMIMONTHLY standard withholding for Married Filing Jointly with adjusted wages of \$1,000?\"}],\"params\":{\"output_types\":[\"multivector\"],\"options\":{\"is_query\":true}}}"

curl https://api.superlinked.com/v1/encode/TomoroAI%2Ftomoro-colqwen3-embed-4b%3Acompact \
  -H "Authorization: Bearer sk-sie-…" \
  -H "Content-Type: application/json" \
  -d "{\"items\":[{\"images\":[{\"data\":\"$page_0\",\"format\":\"png\"}]},{\"images\":[{\"data\":\"$page_1\",\"format\":\"png\"}]},{\"images\":[{\"data\":\"$page_2\",\"format\":\"png\"}]},{\"images\":[{\"data\":\"$page_3\",\"format\":\"png\"}]},{\"images\":[{\"data\":\"$page_4\",\"format\":\"png\"}]},{\"images\":[{\"data\":\"$page_5\",\"format\":\"png\"}]},{\"images\":[{\"data\":\"$page_6\",\"format\":\"png\"}]},{\"images\":[{\"data\":\"$page_7\",\"format\":\"png\"}]}],\"params\":{\"output_types\":[\"multivector\"],\"options\":{\"is_query\":false}}}"

Build the "Visual document search" 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: TomoroAI/tomoro-colqwen3-embed-4b:compact (SIE primitive: /encode). Keep the model id configurable.

Task
- Input: a text query plus one or more document page images.
- Behaviour: return document pages ranked by late-interaction relevance to the query
- Call the selected SIE primitive once per request and map the response into your domain type.

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

Caller-side MaxSim rankingTop 5 of 8

Table header with Married Filing Jointly and Standard withholding columns — 1 · Page 2028.177 MaxSim

2
Page 1726.029 MaxSim
3
Page 1825.723 MaxSim
4
Page 1925.561 MaxSim
5
Page 1525.544 MaxSim

Index a million PDF pages for $200

Serverless API

We host the models; you call one API.

No idle GPUs, pay for what you use
SOC 2 Type 2, US or EU data residency
$50 of credit to start, no card required

Get started

Self-hosted (Kubernetes)

SIE in your clusters, backed by our engineers.

Data stays in your own cloud account
Same models and API as Serverless
Terraform to your cloud in minutes

Talk to us Deploy guide

Run locally

The open-source server on your own machine.

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

Quickstart

Switch to SIE in 5 minutes

import voyageai
from PIL import Image

client = voyageai.Client()
model = "voyage-multimodal-3.5"
page = client.multimodal_embed(
    [[Image.open("page.png")]], model=model, input_type="document",
).embeddings[0]
query = client.multimodal_embed(
    [[question]], model=model, input_type="query",
).embeddings[0]

SIE

from pathlib import Path
from sie_sdk import SIEClient
from sie_sdk.scoring import maxsim

client = SIEClient(api_key="sk-sie-...")
model = "TomoroAI/tomoro-colqwen3-embed-4b:compact"
# Re-index your pages with the new model.
page = client.encode(
    model, [{"images": [Path("page.png").read_bytes()]}], is_query=False,
)[0]["multivector"]
query = client.encode(model, [{"text": question}], is_query=True)[0]["multivector"]
score = maxsim(query, [page])[0]