DocRouter Architecture

DocRouter is a multi-tenant document intelligence platform: ingest documents, run OCR, extract structured data with LLMs, and automate the path into downstream systems. This page describes how the pieces fit together. For deploy and cloud keys, see On-Prem Installation and Platform. For the product walkthrough, see How It Works.

System overview

Layer Role
Frontend (Next.js) Document library, prompts/schemas/tags, PDF review, Flows canvas, admin settings
Backend (FastAPI) REST API: documents, OCR, LLM results, prompts, schemas, tags, flows, webhooks, account config
Workers Async consumers for OCR, LLM extraction, knowledge-base indexing, webhooks, and flow runs
MongoDB App state, versioned prompts/schemas, encrypted credentials, flow definitions and executions, work queues
Blob storage Document binaries and OCR output
LLM providers Inference via LiteLLM (OpenAI, Anthropic, Bedrock, Vertex, Azure, …)
Loading diagram...

Figure 1: DocRouter system architecture — application services, cloud APIs, and OCR/LLM providers.

On-prem, the same services run under Docker Compose or Kubernetes and call managed cloud APIs for storage, OCR, email, and models. See the on-prem architecture diagram.

Request path

  1. UI or SDK/REST client calls the FastAPI backend (org-scoped routes under /v0/orgs/{org_id}/…).
  2. Synchronous work (auth, CRUD, reads) completes in the API process.
  3. Heavy work (OCR, LLM, flow execution, outbound webhooks) is enqueued; workers process messages and update MongoDB (and document state).
  4. Optionally, multi-step workflows are triggered
  5. Clients poll status, subscribe via webhooks, or rely on workflow nodes to send result to target system

Worker pool sizes are configurable per queue (ocr, llm, kb_index, webhook, flow_run).

Document pipeline

The default extraction path is tag- and prompt-driven:

  1. Upload — Document stored; tags select which prompts apply (Tags, Quick Start).
  2. OCR — Org OCR mode runs (e.g. Textract); normalized OCR payload stored for extraction and search (Platform).
  3. LLM extraction — Matching prompts and optional schemas produce structured results via LiteLLM.
  4. Opional Workflow steps, including human-in-the-loop triggers
  5. Export / notify — REST/SDK download, webhooks, or push to ERP.

Statuses progress through states such as ocr_completed → llm_completed. Flows and external workflows extend or replace the “what happens after upload” story.

Automation layer (workflows)

Customers need more than a fixed Upload → OCR → LLM path: branching, schedules, email/drive triggers, agents, and delivery to ERP or review queues. DocRouter supports built-in and external workflows.

Built-in: DocRouter Flows

DocRouter Flows is a first-party visual DAG editor and runtime in the same deployment (no separate Temporal/n8n cluster required).

DocRouter Flows canvas with Gmail trigger, Document Split, OCR, LLM, and HTTP nodes

Triggers include manual, schedule, webhook, chat, poll, and document events. Execution history (inputs, outputs, timing, logs) is available in the UI. Details: Flows and the Flows blog post.

External workflow platforms

When automation lives outside DocRouter, treat DocRouter as the document/OCR/LLM service and orchestrate from elsewhere:

Platform Role
n8n Visual flows with community nodes and SaaS connectors
Power Automate Microsoft cloud flows via the DocRouter custom connector
Temporal Durable coded orchestration
Webhooks + REST API Event-driven or pull-based custom backends

Product webhooks (extraction completed, etc.) are distinct from Flow webhook triggers.

Deployment topology

Mode What you run Cloud / LLM config
Hosted SaaS Nothing — app.docrouter.ai Provided for you (Platform)
Self-hosted Frontend, backend, workers, MongoDB, reverse proxy You supply AWS/GCP/Azure and LLM keys (On-Prem Installation)

Self-host via Docker Compose or Kubernetes. DocRouter does not require AWS Lambda/ECS/EKS for on-prem; it uses cloud APIs (S3, Textract, SES, Bedrock, Vertex, Foundry, etc.) as configured.

Security & credentials

  • Documents and secrets use encryption in transit and at rest where configured for the deployment.
  • Deployment-wide cloud credentials live in MongoDB cloud_config (AWS, GCP, Azure); per-provider LLM keys in encrypted llm_providers.
  • Org and role-based access control scopes documents, prompts, and flows.
  • Audit-oriented logging supports compliance review; enable CloudTrail (and equivalents) in your cloud accounts for infrastructure audit.

Admin UIs: Account → Development for AWS / GCP / Azure setup and LLM Manager. See LLM Configuration.

Integration patterns

How DocRouter typically sits in a larger stack (not alternate product architectures):

DocRouter feeding structured data into an ERP

ERP / ops systems — Extract with prompts or Flows, then POST or sync into ERP, EHR, or databases.

DocRouter as an AI document layer in a larger application stack

AI application layer — Use DocRouter as the document understanding service behind your own product UI and agents.