AI providers: BYOK, hosted & local

Why bundled OCR and single-vendor AI extraction cost 20x too much

• Legacy IDP platforms charge $0.05 to $0.50 per page. Most wrap the same underlying cloud vision models, add a 1,000% token markup, and force users into rigid coordinate-based templates that break on layout changes.
• Bundled single-model tools cannot adapt to document diversity. A clean vector PDF needs a fast deterministic code parser, while a crumpled receipt photo requires a multimodal vision model.
• Sending confidential invoices, patient records, or vendor tax IDs to an unvetted cloud provider creates severe compliance risks under GDPR, HIPAA, and client NDAs.
• Vendor lock-in leaves operations stranded. If a model vendor changes pricing, throttles throughput, or deprecates an endpoint, your document pipeline halts until developers rewrite integrations.

How BYOK document extraction works inside Dynamite Docs

01. Connect your API key in settings

Open Settings > AI & Processing in your workspace. Choose your provider from Google Gemini, Groq, Mistral, OpenAI, Anthropic, OpenRouter, or Custom Endpoint. Paste your secret key and click Verify & Save Key. The server runs an instant health check to discover available models and encrypts the key at rest using AES-256-GCM.

02. Configure active models and fallback routing

Select your default extraction model (such as Gemini 2.0 Flash for sub-second cost efficiency or Claude 3.5 Sonnet for dense financial reasoning). Set a fallback provider to guarantee continuous operation if a primary provider hits rate limits or experiences downtime.

03. Set data privacy and residency policies

Toggle Sensitive Documents Mode to automatically block external providers that permit model training. For strict regulatory compliance, enforce European data residency routes via Mistral AI or route confidential files to the local Ollama companion over a loopback connection.

04. Extract documents with zero platform markup

Upload invoices, bank statements, or receipts. Digital vector PDFs parse locally with code for free. Scans and photos route directly to your connected model. On Pro ($79/mo) and Scale ($199/mo) plans, BYOK extractions cost exactly 0 Processing Entitlements (PE). You pay only raw provider token rates.

Supported providers, endpoints, and architectural controls

Connect any modern multimodal or text LLM. Keys are encrypted with AES-256-GCM and used exclusively to route your authenticated requests:

• Google Gemini: Gemini 2.0 Flash (~$0.0002/page) and Gemini 1.5 Pro (2M token context)
• Groq Cloud: Llama 3.2 Vision at 300-500 tokens/sec on LPU hardware (600ms latency)
• Mistral AI: Pixtral 12B & Large with dedicated EU infrastructure and GDPR compliance
• OpenAI & Anthropic: GPT-4o, GPT-4o mini, and Claude 3.5 Sonnet for complex audit reconciliation
• Cloud Platforms: Azure OpenAI, Amazon Bedrock, Google Vertex AI, and OpenRouter gateways
• OpenAI-Compatible Endpoints: LiteLLM, vLLM, Ollama server, or internal private VPC endpoints
• Local Ollama Companion: 100% on-device extraction on port 8756 with zero cloud transmission
• Zero PE Consumption: Pro and Scale plans include unlimited BYOK runs at 0 PE per page

A realistic example

Realistic routing across a mixed batch of accounting documents:

Request and response examples

Verify and store an encrypted provider key

POST /api/ai/keys

curl -X POST \
  -H "Authorization: Bearer <clerk_token>" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "groq",
    "apiKey": "gsk_a8b9c0d1e2f3...",
    "preferredModel": "llama-3.2-11b-vision-preview"
  }' \
  "https://app.dynamitedocs.com/api/ai/keys"

{
  "success": true,
  "provider": "groq",
  "status": "connected",
  "discoveredModels": [
    "llama-3.2-11b-vision-preview",
    "llama-3.2-90b-vision-preview"
  ],
  "activeModel": "llama-3.2-11b-vision-preview",
  "verifiedAt": "2026-09-15T15:00:00.000Z"
}

Run document extraction using connected BYOK provider

POST /api/v1/infer

curl -X POST \
  -H "Authorization: Bearer dd_live_9f8e7d6c5b4a..." \
  -F "file=@invoice_2026_09.pdf" \
  -F "docType=Invoice" \
  -F "provider=groq" \
  -F "model=llama-3.2-11b-vision-preview" \
  "https://app.dynamitedocs.com/api/v1/infer"

{
  "id": "ext_9821af4e",
  "docType": "Invoice",
  "peCharged": 0,
  "executionTimeMs": 740,
  "provider": "groq",
  "model": "llama-3.2-11b-vision-preview",
  "columns": ["Item", "Qty", "Unit Price", "Total"],
  "rows": [
    ["Server Hosting Q3", "1", "120.00", "120.00"],
    ["Bandwidth Allocation", "5", "15.00", "75.00"]
  ],
  "metadata": {
    "invoiceNumber": "INV-2026-882",
    "subtotal": "195.00",
    "tax": "17.55",
    "total": "212.55"
  }
}

What to validate before relying on it

Verify that the selected provider supports your document modality (multimodal vision vs text) and satisfies active data residency, zero-training, or local loopback policies before processing sensitive files.

Wholesale token economics, encryption architecture, and data residency

Document extraction costs are governed by raw model token consumption, not artificial software tiers. A standard 300 DPI invoice uses approximately 1,200 to 1,600 vision tokens. At Google Cloud published rate of $0.10 per million input tokens for Gemini 2.0 Flash, processing a full invoice page costs approximately $0.0002. Legacy Intelligent Document Processing software charges between $0.05 and $0.50 per page for coordinate OCR, representing a 250x to 2,500x markup on identical underlying compute.

Dynamite Docs Pro ($79/month) and Scale ($199/month) plans eliminate per-page software metering entirely. When you connect an approved provider key, BYOK extractions cost exactly 0 Processing Entitlements (PE). You can process 10,000 or 100,000 pages each month without incurring overage fees or platform lockouts, paying only your raw model token costs directly to Google, Groq, Mistral, OpenAI, or Anthropic. Free and Starter tiers can also connect provider keys within their shared monthly PE budget.

External credentials require strict cryptographic protection. Provider API keys are encrypted at rest with AES-256-GCM using keys managed server-side via BYOK_ENCRYPTION_KEY. Keys are decrypted in memory only for the millisecond window required to sign an outbound HTTPS request to the official provider endpoint. Secrets never land in client-side browser storage, browser network logs, or application crash dumps.

Under standard commercial developer terms from Google Cloud, Groq, Mistral, OpenAI, and Anthropic, data submitted through paid API endpoints is never used to train foundation models. This differs fundamentally from consumer chat apps that retain conversational data. For organizations bound by strict regulatory constraints, Sensitive Documents Mode blocks external providers, while European routes via Mistral keep data within EU boundaries.

When client contracts or healthcare regulations prohibit cloud transmission entirely, the local Ollama companion processes files on-device. The companion runs on port 8756 via an authenticated loopback connection to Ollama on port 11434. Document bytes, raster images, and extracted text never traverse the public internet, continuing to work even during internet outages.

Provider routing security guarantees and boundaries

Does

• Routes requests directly to your verified provider endpoint at wholesale cost with 0% software markup.
• Encrypts API keys at rest with AES-256-GCM and decrypts them in memory only for outbound requests.
• Provides automatic fallback routing to secondary providers during rate limits or service outages.
• Maintains identical schema structure, column names, and validation rules regardless of which AI model runs.
• Permits 100% on-device, air-gapped extraction via the local Ollama companion over loopback port 8756.

Does not

• Does not charge Processing Entitlements (PE) for BYOK runs on Pro or Scale subscription plans.
• Does not expose API keys to browser clients, frontend storage, or audit log exports.
• Does not permit third-party AI vendors to train foundation models on your uploaded documents.
• Does not force rigid coordinate templates that fail when suppliers modify their invoice layout.

Why provider choice and schema locking are non-negotiable

Modern document extraction belongs in a model-agnostic workspace. AI models evolve rapidly, with faster vision architectures and lower pricing launching every quarter. Tying your business operations to a proprietary, single-model tool leaves your workflows vulnerable to price increases and technical stagnation.

Dynamite Docs standardizes the input and output pipeline. Whether an extraction runs through Gemini 2.0 Flash, Groq Llama 3.2, Claude 3.5 Sonnet, or an on-premise Ollama model, your output columns, data validation checks, and spreadsheet formulas remain identical.

Take control of your data residency and operational margins. Connect your provider keys, configure routing policies to match your risk profile, and process documents at true wholesale cost.

Questions, answered

Related integrations

• Document extraction API for PDFs, invoices and OCR
• CSV, JSON, XLSX & document export
• Outbound webhooks

Document workflows it feeds

• GST invoice data extraction without templates
• Accounts payable automation without template setup
• Bank statement extraction to Excel and CSV

Related integration articles

• How to Use Your Own API Key for AI Document Extraction: Complete BYOK Guide
• Invoice Data Extraction Without Templates