The AI Model Garden is Docupath's multi-model AI orchestration engine: the core capability that enables accurate, context-aware document processing without templates or retraining. Instead of routing every document through a single general-purpose model, Docupath dynamically selects, activates, and combines a collection of specialised models based on the document type, structure, and content characteristics detected at ingestion.
The AI Model Garden is the reason Docupath can process a hand-written German invoice and a structured XML purchase order with the same configuration, producing consistent, structured output in both cases.
How It Works
Model Selection and Routing
The AI Model Garden works as an orchestrated workflow rather than a single fixed pipeline. When a document enters the platform, the orchestration layer first detects the document type, then assesses its complexity, language, and quality. Based on that assessment, it decides how to process the document:
Mode selection by complexity: Complex documents - freight invoices or legal invoices, for example - are routed to reasoning models that can interpret context and non-standard structure. Simpler, well-structured documents go through faster, non-reasoning extraction. This keeps throughput high across the queue while giving difficult documents the depth they need.
Multi-modal analysis: Vision reads visual cues that text alone can't - for example, identifying the buyer and seller from a logo or position on the page - and machine learning handles handwritten text.
Deterministic algorithms: Algorithms apply transformation rules and cross-compare values, such as reconciling line items against totals, in a reproducible way.
By selecting and combining the right tools per document - reasoning where it's needed, vision and machine learning for visual and handwritten content, and deterministic algorithms for computation and validation - the Model Garden handles a wide range of document types and qualities under a single configuration.
In-Context Engineering
Rather than retraining models for each customer's document formats, Docupath uses In-Context Engineering: at processing time, the orchestration layer injects customer-defined Instruction Builds directly into the model's context window. This allows the AI Model Garden to adapt to organisation-specific - as well as country-, global-, and trading-party-specific - extraction instructions, field formats, and validation logic without model fine-tuning. Changes to Instruction Builds take effect on the next document processed.
Enrichment and Validation
The AI Model Garden doesn't just extract raw values - it enriches and validates data before surfacing it to the reviewer. Extracted fields are normalised and validated (for example, applying transformation rules and reconciling computed values such as line-item totals), so what reaches the Review interface is enriched, verified information rather than raw model output. Each field is presented alongside its source evidence in the original document, ready for review.
Supported Configurations and Options
Capability | Configuration | Notes |
Document type routing | Automatic | Based on classification at ingestion |
Model activation | Dynamic per document type | No manual model selection required |
In-Context Engineering | Via Instruction Builds | Applied at processing time, no retraining |
Types of Models
Model Type | Function |
Vision | OCR, handwriting recognition, layout parsing, image quality assessment |
Language | Semantic extraction, entity recognition, relationship mapping, context inference |
Layout | Structure-independent parsing, table detection, field localisation |
Domain-specific | Regulated field types (VAT, IBAN, HS codes, currency, dates) |
Other Technical Specifications
Metric | Value |
Supported input languages | Multi-language (coverage varies by document type) |
Processing approach | Asynchronous; results available after extraction completes |
In-Context Engineering latency | Zero; applied at runtime, no deployment cycle |
Model updates | Managed by Docupath; transparent to customers |