NDI · Multimodal Perception Layer
The perception layer for AI agents.
Feed NDI raw multimodal data — documents, audio, images and video — and get back bite-sized, agent-ready Markdown and JSON, grounded to the exact page, cell or timestamp. Best-in-class accuracy, whether you're inside an agent's tool loop or clearing a million-file backlog.
POST /parse · /ground · /categorize · /extract · sync & async
Methods
Four methods. Any document workflow.
Every method is one REST endpoint with a sync and an async variant — same request body, same job contract, safe retries with idempotency keys.
Works where your agents work
Every method is available over REST and as an MCP server — plug NDI into Claude, Cursor or any MCP-compatible agent.
Quickstart
Your first parse call in one request.
One endpoint, one header. The sync call blocks until the result is ready; the async variant returns a job_id you can poll — same body, same contract, safe retries with idempotency keys.
ndi — quickstart
POST /parse · sync or async · durable by design
$ curl https://ndi-api.nace.ai/api/v1/parse \
-H "X-API-Key: $NDI_API_KEY" \
-d '{"file": {"source_url": "https://…/annual-report.pdf",
"file_name": "annual-report.pdf"}}'// 200 OK · "status": "succeeded"
"result": {
"markdown_url": "https://…",
"page_count": 96,
"char_count": 184223
}25+ formats
If your team works with it, NDI can digest it.
Native documents, scans, spreadsheets, presentations, emails, images, diagrams — one agent, one API, no format-specific plumbing. The format is inferred from the file extension, and unsupported files fail fast before a job ever starts.
Uploads up to 200 MB per file, or point NDI at any https:// or s3:// URL up to 1 GiB. If a format matters to you and isn't listed, talk to us — the ingestion layer is extensible.
How it works
An agent that reads like your best analyst.
NDI doesn't just OCR pixels — it reasons over structure, checks its own work, and never returns a value it can't point to.
Understand the layout
High-fidelity parsing reconstructs the document's real structure — headings, tables, figures, multi-column flow — into page-marked markdown any model can consume.
# coastal survey
## shoreline
| section | conf |
| ————— | ——— |
| shoreline | 0.98 |
<!-- page 4 -->
Reason & ground
The agent extracts what you asked for, then grounds each item against the source with verbatim evidence and bounding boxes. Ambiguity triggers region-level re-reads, not guesses.
Deliver, durably
Every job runs as a typed, durable workflow with automatic retries and full run visibility. Fire-and-forget async, or synchronous for short jobs — same contract either way.
Evals · OmniDocBench v1.5
State of the art, measured in public.
NaceOCR-V2 — the parsing model behind NDI — leads every headline metric on OmniDocBench v1.5, a public benchmark of 1,651 real document pages, against systems many times its size.
OmniDocBench's headline leaderboard score — 6.1 points ahead of the next best system.
Best overall score on the benchmark — 2.1× lower than the next best model.
+13.2 points over the next best system — tables parsed with structure and text intact.
SOTA accuracy at 2.6 seconds per page — 3–5× faster than the deep-extraction systems it beats.
Composite ↑ (higher is better)
Overall Edit distance ↓ (lower is better)
| Model | Latency / Page ↓ | Composite ↑ | Overall Edit ↓ | CDM ↑ | Table TEDS ↑ | Reading-Order ↓ |
|---|---|---|---|---|---|---|
| NaceOCR-V2 | 2.6s | 94.8 | 0.081 | 0.969 | 91.2% | 0.123 |
| Mistral OCR 4 | 2.6s | 88.7 | 0.224 | 0.957 | 76.7% | 0.146 |
| Datalab Marker (Chandra) | 7.3s | 87.9 | 0.231 | 0.952 | 74.9% | 0.148 |
| Reducto Deep Extract | 12.5s | 84.9 | 0.167 | 0.903 | 78.0% | 0.197 |
| Reducto Parse | 3.1s | 61.2 | 0.381 | 0.330 | 66.3% | 0.238 |
| LlamaParse | 8.4s | 34.5 | 0.648 | 0.312 | 30.6% | 0.459 |
OmniDocBench v1.5 — 1,651 public benchmark pages. Composite is the leaderboard headline score, ((1 − Text Edit) × 100 + Table TEDS + Formula CDM) / 3 (higher is better). Overall Edit averages text, formula, table and reading-order edit distances (lower is better). CDM = visual formula matching; Table TEDS = table structure + extracted text (higher is better). Latency is median seconds per page, measured on our synthetic benchmark under identical conditions.
Industries
Where the documents are messy and the stakes are high.
Every industry has its own flavor of messy — handwritten margins, skewed scans, phone photos, redaction bars. The demos below are real files run through the live API: hover a field and the lens moves to the exact spot on the page NDI grounded it to.
PDF · NF-e invoice · 1 page · parsed in 6.1s · live API output
- Ingest general ledgers up to 1 billion lines
- Tie-outs grounded from schedule back to source evidence
- Handwritten tie-out notes and stamps on scanned support
- Full PBC folders — mixed formats, thousands of pages, async
Deployment · built for privacy
Your documents never have to leave your walls.
Run NDI wherever your risk profile demands — the API contract is identical everywhere.
Self-hosted
Deploy on your virtual cloud, edge or on-premises. Maximum customization and control — your data stays within your walls.
Managed service
A fully managed cloud service that handles deployment, scaling and maintenance — ease of use without the ops burden.
Hybrid
Keep sensitive workloads in-house and burst to managed capacity for scale — combine both models per workload.
FAQ
Questions, answered.
Make your first parse call in minutes.
One API. 25+ formats. Grounded output your team can actually trust.