The archive
that organises itself
on its own.

The ingestion layer presents consistent interfaces to four completely different input modes. A scan on a USB stick, a drag-and-drop folder on a PC, an email forwarded by a supplier, a photo sent from a smartphone: all four converge into the same pipeline.

The orchestration layer applies the business logic: detects duplicates, decides who can see what, routes the classification, notifies the recipients. It is the part that knows the company's org chart.

The AI layer — Forge AI — is a proprietary multi-provider engine that abstracts the underlying language model. The rest of the system does not know whether it is talking to Gemini, Claude, OpenAI or a local model: it talks to a stable interface.

The deterministic stages (1-2, 4-6) take less than three seconds overall and run entirely on the local server. The dominant latency is the call to the external AI provider, which varies between eight and twelve seconds for an average document. Optimisations such as lighter models, batching and caching reduce this time for high-volume pipelines.

Time per document: manual vs. Vdoc

A secretary who receives an invoice by email spends on average three to four minutes between: opening it, renaming the file, choosing a folder, entering the metadata in the management system, forwarding it to the person responsible. Over 80 documents a month, that is roughly five hours per month — sixty hours a year. Vdoc reduces the human work to merely verifying the already-classified document.

Distribution of manual work

Cumulative annual cost

Estimates based on times measured in pilot installations and average commercial AI provider rates for documents of 1 to 5 pages. Higher volumes → decreasing marginal costs.

A company can configure multiple instances of each channel — several monitored email addresses, several registered PC workstations, several authorised Telegram users — assigning each its own rules and permissions.

Forge AI is a provider-agnostic abstraction layer built around three principles: versioned prompts, validated outputs, trackable costs. Vdoc uses it to extract structured metadata from documents without coupling to a specific AI vendor.

Four providers, a single interface

What Forge AI does, beyond the call

• Versioned prompts. The instructions sent to the AI are saved as immutable versions identified by a code. Improving a prompt does not break the existing ones: the old version stays as a reference, the new one enters production.
• Structured validation. Every AI response is verified against a schema. If the AI returns malformed JSON or a field is missing, the system automatically retries the request or reports the error without corrupting the archive.
• Spend quotas and caps. Monthly limits in tokens and euros, configurable per installation. The system warns before exhaustion and blocks new calls when the cap is reached, avoiding billing surprises.
• Intelligent throttling. When channels push documents in rapid sequence, Forge AI automatically adjusts the call rate to respect provider limits, with no HTTP 429 errors visible to the user.
• Multi-step chains. For cases where a single call is not enough (for example: extract the metadata, then verify consistency), Forge AI orchestrates call chains with automatic output propagation.
• Complete tracing. Every call is logged: provider, model, tokens, estimated cost, latency, outcome. The dashboard shows real-time consumption by day, by channel, by document type.
• Encrypted API keys. Provider credentials are never in plain text on the server. Fernet encryption with a local master key and strict file-system permissions.

from forge_ai import Gateway, ChatRequest

gw = Gateway(config_path="/etc/vdoc/forge_ai.toml")

risposta = gw.chat(ChatRequest(
    prompt_codice="vdoc.classifica_documento",
    prompt_versione="v3",
    documenti=[documento_pdf],
    output_schema=SchemaClassificazione,
    metadata={"canale": "email", "mittente": "fornitore@acme.it"},
))

# risposta.output è già validato contro lo schema
# risposta.costo_stimato_usd, risposta.token_*, risposta.latenza_ms
# tutto è loggato automaticamente in forge_ai.chiamata

The overlap is intentional. Vdoc does not replace Paperless-ngx: it is its main consumer. Whoever administers the system can access the Paperless interface directly for advanced operations — manual metadata editing, complex search, bulk export — without having to wait for dedicated features in the Vdoc dashboard.

On first access a guided procedure walks the administrator through the minimum necessary configuration. Afterwards, the panel stays open, with a list of optional configurations still to be defined.

Internal segregation

Users see only the documents of their groups. The segregation is enforced by the native Paperless permission engine and cannot be bypassed from the user side. The system administrator is the only one with full technical visibility, as per standard practice.

Operational log

Every processed document leaves a record: who sent it, when, with what outcome, with what metadata. The event archive can be consulted from the Vdoc dashboard and is kept for the configured period.

Hosting: VPS or on-premise

The system is designed to run on a single modest machine — typically a VPS with 4 GB of RAM and 2 vCPUs in a European data centre is sufficient for companies up to about 500 documents per month. For higher volumes, or for internal data-residency requirements, the system installs identically on a physical server hosted in-house.

• It is not a management system. It does not record accounting entries, does not issue invoices, does not manage active deadlines. It files and classifies documents that come from other systems.
• It is not a PEC or an interchange system. Filing electronic invoices goes through ordinary channels; Vdoc can receive copies via email after they have been acquired.
• It does not enforce approval workflows. A filed document is filed; there are no intermediate states like 'awaiting signature' managed by the system. Approval takes place in the existing company processes.
• It does not delete automatically by regulatory deadline. Fiscal and legal retention rules are the company's responsibility. The system offers only optional retention and protections against accidental deletion, but it does not decide on its own what to eliminate.
• It does not interpret content beyond classification. The automatic summary is an aid to search, not a substitute reading of the document. Decisions based on content require human reading.