AI Model Strategy

The platform should support multiple AI providers behind a routing layer. The business workflow should not depend directly on one provider API.

Provider comparison

Provider	Strengths	Best use
OpenAI GPT	Strong structured extraction, reliable JSON, good reasoning over mixed documents.	Rental contracts, complex bills, ambiguous documents.
Anthropic Claude	Strong long-context reading and careful language understanding.	Long contracts, dense legal documents, multi-page schedules.
Google Gemini	Strong multimodal handling and broad document understanding.	Scanned documents, image-heavy bills, mixed visual layouts.
DeepSeek	Cost-effective text reasoning for simpler extraction tasks.	Invoices, straightforward schedules, second-pass validation.

Routing strategy by document type

Document type	Primary model	Fallback
Rental contracts	OpenAI GPT or Claude	Gemini for image-heavy scans
Utility contracts	OpenAI GPT	Claude
Utility bills	Gemini or OpenAI GPT	DeepSeek for normalized text
Bank transaction fallback	OpenAI GPT	DeepSeek
Invoices	DeepSeek or OpenAI GPT	Gemini for scanned invoices
Tax documents	OpenAI GPT	Claude
Condominium schedules	Claude	OpenAI GPT
Insurance contracts	Claude or OpenAI GPT	Gemini
Property ownership documents	Claude	OpenAI GPT

Fallback logic

Fallback should be triggered when:

• JSON is invalid
• classification confidence is low
• required fields are missing
• OCR output is poor
• model returns unsupported values
• backend validation fails for format-sensitive fields

Prompt versioning

Prompts should be versioned by document type. A saved extraction job should record:

• provider
• model
• prompt version
• schema version
• token usage
• latency
• confidence summary

Practical recommendation

Start with one high-accuracy model for the first production workflows. Add routing after enough extraction logs exist to identify which document types are expensive or error-prone.