Give every workflow decision its own scored answer

SIE answers your named questions together, with a score for each choice or yes-or-no check. Run the open model beside your data; your application decides what happens next.

Run the example Self-host from main

Invoice review · INV-2026-6576

Vendor: Kestrel Supply
Invoice total: $11,087
Purchase order: $11,087
Delivery: 100 units, accepted
Prior invoice IDs: INV-2026-6576INV-2026-5581INV-2026-5526

Recorded SIE answers

Matches the order and deliveryYes 99.8% model score

Previously submitted invoiceYes 94.6% model score

Your application rule

duplicate = yesReview before paying

Name the checks your workflow needs

One request, several answers

Keep the next action in your app

Read each named score, then apply your review and action rules.

Check the conditions behind a single label

Archived source · CVE-2023-6826

The E2Pdf plugin for WordPress is vulnerable to arbitrary file uploads due to insufficient file type validation on the 'import_action' function in versions up to, and including, 1.20.25. This makes it possible for authenticated attackers with a role that the administrator previously granted access to the plugin, to upload arbitrary files on the affected site's server which may make remote code execution possible.

Read the NVD record

Separate named answers

Attack vector: ✓ remotely over the network
Can the attacker exploit it without a login?: ✓ No

Recorded GLiClass Large answers. Network reachability and account access are independent checks.

See accuracy for each named check

Correct answers on 160 CVE records

% higher is better

Attack vector

Attack vector 91.9%

Weakness class

Weakness class 88.8%

Attacker needs no login

Attacker needs no login 71.9%

Every check on one record

Every check on one record 60.0%

160 held-out records

Recorded 24–25 September 2026. GLiClass Large v1.0, three named questions per December 2023 CVE record. Each record ran once.

Every check was right on 60.0% of complete records. The 2026 invoice is a separate workflow illustration.

How we tested Run the evaluation Recorded outputs

Name the checks, read each answer

Record + named question groups

result = client.extract(
    "knowledgator/gliclass-large-v1.0",
    {"text": json.dumps({
        "invoice": {"id": "INV-2026-6576",
            "total_usd": 11087, ...},
        "purchase_order": {"total_usd": 11087, ...},
        "vendor_history": {"prior_invoice_ids": [
            "INV-2026-6576", ...]},
        ...
    })},
    options={"label_groups": {
        "matches_order": [...],
        "duplicate": [...],
        ...
    }, "group_encoding": "separate"},
)

Abbreviated saved request. The complete request includes every invoice field and all five question groups.

View complete request

import json
from sie_sdk import SIEClient

record = "{\"delivery\": {\"condition\": \"accepted\", \"date\": \"2026-03-07\", \"received_qty\": 100}, \"invoice\": {\"currency\": \"USD\", \"id\": \"INV-2026-6576\", \"lines\": [{\"qty\": 100, \"sku\": \"SKU-286\", \"total_usd\": 11087.0, \"unit_usd\": 110.87}], \"total_usd\": 11087.0, \"vendor\": \"Kestrel Supply\"}, \"payment\": {\"days_until_due\": 44, \"discount_expires_in_days\": null, \"early_payment_discount_pct\": null, \"status\": \"scheduled\", \"terms\": \"net 30\"}, \"purchase_order\": {\"freight_terms\": \"freight prepaid by vendor, not separately billable\", \"id\": \"PO-8988\", \"lines\": [{\"qty\": 100, \"unit_usd\": 110.87}], \"total_usd\": 11087.0}, \"vendor_history\": {\"disputes_12m\": 0, \"invoices_12m\": 52, \"prior_invoice_ids\": [\"INV-2026-6576\", \"INV-2026-5581\", \"INV-2026-5526\"]}}"
questions = json.loads("{\"discrepancy_severity\":[\"None: everything reconciles.\",\"Trivial: rounding or a cosmetic difference.\",\"Moderate: a real difference worth confirming.\",\"Material: a large or unexplained difference.\"],\"disposition\":[\"approve: Matches the order and delivery; approve for payment.\",\"hold: Something needs confirming before payment; hold pending clarification.\",\"manual review: A human in finance must review the discrepancy.\",\"reject: Should not be paid: duplicate, unauthorised or materially wrong.\"],\"duplicate\":[\"This invoice appears to duplicate an invoice already submitted.\",\"Not the case: this invoice appears to duplicate an invoice already submitted.\"],\"matches_order\":[\"Line items, quantities and amounts reconcile.\",\"There is a discrepancy against the order or the delivery.\"],\"urgency\":[\"No time pressure; can wait indefinitely.\",\"Routine; handle within the normal queue.\",\"Elevated; should be handled within the same week.\",\"Critical; requires action within the same day.\"]}")
with SIEClient("http://localhost:8080") as client:
    result = client.extract(
        "knowledgator/gliclass-large-v1.0", {"text": record},
        options={
            "label_groups": questions,
            "group_encoding": "separate",
            "overflow_policy": "truncate_text",
        },
    )
print(result["data"])

Output

{
  "matches_order": {
    "answer": true,
    "model_score": 0.9976
  },
  "duplicate": {
    "answer": true,
    "model_score": 0.9462
  }
}

Two shown answers from the recorded five-question request
Your app applies its own review and action rules

Switch to SIE in 5 minutes

Full runnable recipe

reply = anthropic.messages.create(
    model="claude-sonnet-5", max_tokens=400,
    system="Answer each named question about the record.",
    messages=[{"role": "user", "content":
        json.dumps({"record": record, "questions": questions})}],
)
answers = json.loads(reply.content[0].text)

SIE

# Keep the same record and named question groups.
result = client.extract(
    "knowledgator/gliclass-large-v1.0",
    {"text": record},
    options={
        "label_groups": questions,
        "group_encoding": "separate",
        "overflow_policy": "truncate_text",
    },
)
answers = result["data"]

Source attribution and license

This product uses data from the NVD API but is not endorsed or certified by the NVD.

MITRE hereby grants you a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare derivative works of, publicly display, publicly perform, sublicense, and distribute Common Vulnerabilities and Exposures (CVE). Any copy you make for such purposes is authorized provided that you reproduce MITRE's copyright designation and this license in any such copy. (CVE Program Terms of Use, https://www.cve.org/Legal/TermsOfUse)

Keep your records and decisions in your cloud

Self-host with K8s

Easy & scalable deployment in your own cloud.

Terraform to your cloud in minutes
Apache-2.0, same engine as Cloud
Scales to zero, no bill between jobs
Per-tenant pools, no noisy neighbors

Agent prompt

Deploy SIE to our AWS account with the superlinked/sie/aws Terraform module. Docs: superlinked.com/docs/deployment Serve knowledgator/gliclass-large-v1.0 using the public model configurations.

Deploy SIE to our GCP project with the superlinked/sie/google Terraform module. Docs: superlinked.com/docs/deployment Serve knowledgator/gliclass-large-v1.0 using the public model configurations.

Deploy SIE to our Azure AKS cluster via helm install. Requirements: superlinked.com/docs/deployment Serve knowledgator/gliclass-large-v1.0 using the public model configurations.

Deploy guide

Run locally

Score named checks on your NVIDIA GPU.

NVIDIA GPU
One container, no cluster
GLiClass Large v1.0, fully offline
Same SDK and IDs, no code changes

docker run --gpus all -p 8080:8080 \
  ghcr.io/superlinked/sie-server:v0.9.0-cuda12-default \
  serve --host 0.0.0.0 --port 8080 \
  --models knowledgator/gliclass-large-v1.0

Quickstart