Multi-Modal Guardrails

Palveron supports multi-modal governance verification. Send text, images, audio, documents, and code in a single request — the gateway auto-detects the content type and routes to the appropriate analysis provider.

Supported Content Types

Content type	Provider	Analysis
Text	GPT-4o-mini	PII detection, prompt injection, policy evaluation
Image	Llama Guard Vision	NSFW detection, content policy, OCR + text analysis
Audio	Whisper	Speech-to-text transcription → text analysis pipeline
Document	Text extraction	PDF/DOCX text extraction → text analysis pipeline
Code	Secret Scanner	API keys, AWS credentials, GitHub tokens, private keys
Video	Stub (Phase 2)	Frame extraction → image analysis pipeline

Sending Attachments

Attachments are sent base64-encoded with a MIME type. The SDK handles encoding automatically when using the file helpers.

TypeScript

// From file path
const result = await palveron.verifyWithFile(
  'Analyze this document for compliance issues',
  './contract.pdf'
);

// From base64
const result = await palveron.verify({
  prompt: 'Check this image for sensitive content',
  attachments: [{
    contentType: 'image/png',
    data: base64EncodedImage,
    filename: 'screenshot.png',
    metadata: { source: 'user-upload' },
  }],
});

Python

from palveron import Palveron, VerifyRequest, Attachment

client = Palveron(api_key="pv_live_...")

# From file
result = client.verify_file("Check this", "/path/to/image.png")

# Programmatic
result = client.verify(VerifyRequest(
    prompt="Analyze this code for secrets",
    attachments=[Attachment.from_file("app.py")],
))

# From raw bytes
att = Attachment.from_bytes(image_bytes, "image/jpeg", "photo.jpg")

Go

att, _ := palveron.AttachmentFromFile("/path/to/contract.pdf")
result, _ := client.Verify(ctx, &palveron.VerifyRequest{
    Prompt:      "Check this document",
    Attachments: []palveron.Attachment{*att},
})

cURL

BASE64=$(base64 -i photo.jpg)

curl -X POST https://gateway.palveron.com/api/v1/verify \
  -H "Authorization: Bearer pv_live_..." \
  -H "Content-Type: application/json" \
  -d "{
    \"prompt\": \"Check this image\",
    \"attachments\": [{
      \"content_type\": \"image/jpeg\",
      \"data\": \"$BASE64\",
      \"filename\": \"photo.jpg\"
    }]
  }"

Response

Multi-modal responses include content_type and findings:

{
  "decision": "BLOCKED",
  "reason": "Secret detected in code",
  "content_type": "code",
  "findings": [
    {
      "risk": "CRITICAL",
      "category": "aws_credentials",
      "description": "AWS Access Key ID detected: AKIA...",
      "confidence": 0.99
    }
  ],
  "trace_id": "clx...",
  "integrity_hash": "sha256:abc...",
  "should_anchor": true
}

Sensitivity Levels

Configure sensitivity per project under Settings → Multi-Modal:

Level	Blocks on
Low	CRITICAL findings only
Medium (default)	HIGH + CRITICAL
High	MEDIUM + HIGH + CRITICAL

Code Secret Scanning

The code provider scans for hardcoded secrets in any programming language:

Pattern	Example
AWS Access Keys	`AKIAIOSFODNN7EXAMPLE`
AWS Secret Keys	`wJalrXUtnFEMI/K7MDENG/bPxRfiCY`
GitHub Tokens	`ghp_xxxxxxxxxxxxxxxxxxxxxxxxxxxx`
Stripe Keys	`sk_live_...`, `sk_test_...`
Generic API keys	`sk-proj-...`, `Bearer ...`
Private Keys	`-----BEGIN RSA PRIVATE KEY-----`

Configuration

Multi-Modal can be configured in three modes under Settings → Multi-Modal:

Mode	Description
Managed	Palveron-hosted analysis providers. No setup required.
Self-Hosted	Your own Llama Guard / Whisper instances. Full data sovereignty.
Disabled	Multi-modal analysis off. Only text is verified.

Architecture

verify({ prompt, attachments })
  → Content-type detection (auto: OpenAI / Anthropic / Palveron format)
  → Provider routing: Text → GPT-4o-mini
                      Image → Llama Guard Vision
                      Audio → Whisper → text pipeline
                      Code → Secret Scanner
  → Merge findings
  → Apply sensitivity filter
  → Decision (ALLOWED / BLOCKED / MODIFIED)
  → Trace + Flare anchoring

All content types pass through the same Flare blockchain anchoring pipeline. contentType and contentMetadata.findings are stored on every trace for audit purposes.

On this page