Skip to main content
Nebula

Charts, tables, and business files become structured data your AI can actually answer from.

Free to try with your own file. Results in minutes.

99%accuracy on machine-typed documents
97%accuracy on handwritten numbers
92%accuracy on complex graphs and tables

Accuracy = the share of questions a downstream LLM (GPT-5.4-mini or Gemini 3 Flash) answers correctly from Nebula's converted output, matching how a human reads the original document. Full methodology report coming soon.

Input document
100%
M&A Annual Report 2025 — Global deal activity
Nebula output

Exhibit 2

Globally, M&A activity increased 12 percent in 2024.

Announced global M&A deals, 2015–24¹

<!-- page-background: white -->

Chart: Deal Value ($ billion) - Bar Chart

{
  "type": "bar",
  "data": {
    "labels": ["2015", "2016", "2017", "2018", "2019", "2020", "2021", "2022", "2023", "2024"],
    "datasets": [
      {
        "label": "Deal value, $ billion",
        "data": [4500, 3800, 3600, 4200, 4000, 3700, 5800, 3600, 3100, 3500],
        "backgroundColor": "#002A6D"
      }
    ]
  },
  "options": {
    "scales": {
      "y": {
        "title": {
          "display": true,
          "text": "Deal value, $ billion"
        },
        "beginAtZero": true,
        "ticks": {
          "stepSize": 1000
        }
      },
      "x": {
        "title": {
          "display": true,
          "text": ""
        }
      }
    },
    "plugins": {
      "title": {
        "display": false
      }
    }
  }
}

Chart: Deals (thousand) - Line Chart

{
  "type": "line",
  "data": {
    "labels": ["2015", "2016", "2017", "2018", "2019", "2020", "2021", "2022", "2023", "2024"],
    "datasets": [
      {
        "label": "Deals, thousand",
        "data": [8.5, 8.6, 8.7, 8.8, 8.5, 8.4, 12.5, 9.8, 7.2, 7.8],
        "borderColor": "#007BFF",
        "fill": false,
        "pointRadius": 4,
        "pointBackgroundColor": "#007BFF"
      }
    ]
  },
  "options": {
    "scales": {
      "y": {
        "title": {
          "display": true,
          "text": "Deals, thousand"
        },
        "beginAtZero": true,
        "ticks": {
          "stepSize": 2
        }
      },
      "x": {
        "title": {
          "display": true,
          "text": ""
        }
      }
    },
    "plugins": {
      "title": {
        "display": false
      }
    }
  }
}

(World map icon)

¹ Announced deals not withdrawn or canceled; deal value >$25 million.
Source: Dealogic; McKinsey M&A Insights

McKinsey & Company

¹⁴ Anat Alon-Beck, “New FTC rules tighten the screws on private equity in mergers,” United States Law Week, October 21, 2024.
¹⁵ Ryan Phair, Michael Murray, and Catherine Kordestani, “Dealmakers face a time-consuming filing burden under new rules,” United States Law Week, October 24, 2024.

How many M&A deals in 2021?

~12,500 deals

AI can answer correctly

from the source

An annual M&A outlook page with narrative text, a bar chart, and a line chart. Nebula extracts the page text and returns the chart data as Chart.js-compatible JSON inline.

The problem

Looks right is not the same as your AI can use it.

Most OCR optimizes for clean-looking text. But a perfectly transcribed table that loses its rows and columns is unanswerable. Nebula optimizes for the metric that matters: whether a downstream model can answer a question correctly from the result, just as a human would from the original.

Typical OCR: looks clean, unanswerable
Segment Cloud 12 19 27 41 Services 8 9 11 14
(rows and columns flattened)

Ask "Cloud revenue in FY24?" and the model has to guess: the answer (27) is buried in a flat run of numbers.

Nebula: structured, answerable
| Segment  | FY22 | FY23 | FY24 | FY25 |
| -------- | ---: | ---: | ---: | ---: |
| Cloud    |   12 |   19 |   27 |   41 |
| Services |    8 |    9 |   11 |   14 |

Ask the same question and a downstream LLM answers correctly: FY24 Cloud = 27.

Accuracy = the share of questions a downstream LLM (GPT-5.4-mini or Gemini 3 Flash) answers correctly from Nebula's converted output, matching how a human reads the original. Full methodology report coming soon.

The hard stuff

Charts to Chart.js JSON. Tables that keep their shape.

The structure that makes a document useful survives the parse, so your AI gets data instead of flattened text.

Charts become data

Bar, line, and pie charts are transcribed into structured Chart.js JSON: labels, series, values, and axis units. Numbers stay numbers.

Tables keep structure

Rows, columns, merged cells, and financial-statement hierarchy survive the parse, with footnotes preserved instead of dropped.

Three ways to adopt Nebula

Use the platform, connect the API, or deploy your own model.

Choose the path that matches your workflow: a web platform for teams, an API for pipelines, or a private fine-tuned deployment for enterprise AI workloads.

Nebula app

Sign in, drag in a PDF, an image deck, or a whole folder, watch pages process, and download chart-aware Markdown and JSON. No code required.

  • Upload and go, no setup
  • Batch and folder uploads
  • Download Markdown or JSON

Nebula API

Async REST with presigned uploads, batch and folder jobs, and Markdown or JSON output for your RAG and agent pipelines.

  • Async REST, presigned uploads
  • Batch and folder jobs
  • Markdown, JSON, or both
Create a job
POST /v1/jobs
{
  "model_id": "nebula",
  "options": { "response_format": "both" }
}

Local fine-tuned deployment

Deploy a private model in your environment for sensitive, repeatable enterprise workloads that should run on systems you own.

  • Scoped to your architecture
  • Fine-tuned on your document domain
  • Designed for on-premises or private environments
Enterprise model enablement

Beyond the API: own and tune the models behind your AI.

For enterprises standardizing on AI, Nebula is also how you provision, route, and fine-tune the models underneath. This is the third way to adopt Nebula, beyond the web platform and API.

Model routing

Route each task to the best frontier or fine-tuned model through a single integration point, with no vendor lock-in.

Local & private LLM provisioning

Provision private, locally-deployed LLMs inside your environment for sensitive workloads, scoped to your architecture.

Domain fine-tuning

Fine-tune models on your own documents and domain, so everyday workflows run on systems you own and control.

Evaluation & integration harnesses

Wire models into your systems and keep quality measurable with evaluation harnesses built around answerability.

Deployment model, SLAs, and latency are scoped per customer architecture and commercial terms.

See our evaluation methodology
Models & trust

The right model for every document, with enterprise controls.

Nebula routes to leading frontier vision-language models plus our own fine-tuned VLMs, choosing the best for each document. Documents are processed through enterprise model providers under their enterprise terms, kept isolated per tenant, and never used to train models.

Best model per document
Nebula routes to leading frontier vision-language models plus our own fine-tuned VLMs, picking the strongest model for each document and layout.
Per-tenant isolation
Source uploads, intermediate artifacts, and final output are isolated per tenant, with per-customer API-key authentication.
Not used to train models
Your documents are processed via enterprise model providers under their enterprise terms and are not used to train models.
Per-job audit trail
Per-customer API keys, plus per-job timing, status, and page-level outcomes for security and finance review.
Security and trust
Enterprise controls on every job.
  • Per-tenant data isolation
  • Per-customer API-key authentication
  • Per-job audit trail
  • Documents are not used to train models
  • Processed via enterprise model providers under their enterprise terms

ISO 27001 certified, with SOC 2 and GDPR compliance in progress.

Use cases

One layer for many document workflows.

Use Nebula wherever business knowledge needs to move from visual documents into AI systems: RAG, agents, analytics, and downstream LLM pipelines.

01

Japanese financial documents

Annual reports, governance materials, IR releases, and disclosure filings become structured AI inputs without losing Japanese layout.

02

Statements & invoices

Issuer-specific statement and back-office layouts become normalized line items for accounting and reconciliation agents.

03

Board decks & presentations

Slide structure, narrative, tables, and chart series are extracted from PowerPoint exports without flattening the visual context.

04

Data rooms & RAG ingestion

Large enterprise PDF estates are prepared for retrieval, summarization, diligence, and agentic workflows.

05

Legal & regulatory corpora

Long Japanese legal PDFs are ingested with footnotes and citations preserved in reading order.

06

Construction, logistics, retail

Inspection sheets, registries, delivery notes, and forms are structured consistently across industries.

FAQ

Frequently asked questions about Nebula.

What Nebula is, how it processes documents, where the data lives, and how to integrate.

Try Nebula

Give your AI documents it can answer from.

Upload a document and get chart-aware Markdown plus structured JSON in minutes, or call the API from your pipeline.

Best frontier and fine-tuned VLMs, routed per document
Markdown + structured JSON for RAG and agents
Tuned for Japanese financial, legal, and business documents