smallDat — Document Analysis Platform

Why smallDat

Business value at a glance

Capture expert know-how once. Run it consistently, with any model, and prove what happened.

Build once, reuse forever

Analysts capture know-how as templates
Every run is consistent & repeatable

Any LLM, no lock-in

OpenAI, Claude & Gemini side by side
Swap models without touching the flow

Answers in real time

Results stream as each step finishes
No black-box waiting

No-code visual builder

Drag-and-drop analysis graphs
Domain experts build, not engineers

Audit & history built in

Every run saved and re-openable
Compare, export, prove what happened

Govern who sees what

Role-based: Maker / Consumer / Admin
Multi-team, multi-tenant ready

How it works

Two roles, one pipeline

MAKER · builds the workflow

Designs reusable analysis templates
Configures models, activities & documents
Manages users, groups & access

A curated catalog of ready-to-run analyses the whole org can trust.

➜

CONSUMER · runs the workflow

Picks a template, uploads documents
Runs it and watches results stream
Browses history, chats, exports

Structured answers in minutes — no prompt engineering required.

Maker

Design the analyses — once

Design reusable analysis templates

The visual Flow Designer turns an analysis into a graph of steps — built once, run by anyone.

Drag activities onto a canvas and wire them into a flow
Each node calls an LLM with its own prompt & parameters
Chain steps: summarise → extract → format → combine
Save as a template in the Consumer catalog

Business value

Capture expert know-how as a repeatable asset — not a one-off chat.

Configure the building blocks

Makers control the models, activities and document inputs every template draws on.

Models — register OpenAI, Claude or Gemini endpoints + keys
Activities — reusable steps with typed prompts & inputs
Document definitions — what files a flow expects
Override model or prompt per node, without code

Business value

One place to manage providers, cost and prompts across all flows.

Govern users & access

Role-based groups decide who can build and who can run.

Create groups with roles: Consumer, Maker, Administrator
Assign users to groups; gate access to app areas
Makers curate templates; Consumers only run them
Multi-team / multi-tenant friendly

Business value

Safe self-service: experts build, everyone else runs — with guardrails.

Consumer

Get answers — in minutes

Start from a template, drop in documents

No setup, no prompts — pick a ready-made analysis and feed it your files.

Browse the catalog of Maker-published templates
Start a session in one click
Drag & drop PDFs (and more) as inputs
Inputs validated before you run

Business value

Zero learning curve — value from the first document.

Run it and watch results stream

The workflow engine executes each step and pushes results live over WebSocket.

One Run button kicks off the whole graph
Per-step results appear as they complete
See timing and status for every activity
Stop or re-run at any time

Business value

Answers as they happen — not a spinner and a long wait.

Browse history, chat & export

Every run is saved, comparable, and ready to share.

Re-open any past run and drill into single steps
Chat with an LLM over a run's results
Export runs (JSON) and outputs (PDF / Word / Excel)
Import a shared run to inspect it

Business value

A durable, auditable record — reuse and share, don't repeat.

Behind the scenes

Architecture & engine

Async Python core, a graph-based workflow engine, and a real-time React canvas — typed end to end with Pydantic v2.

Workflow engine

LangGraph StateGraph built per template
Nodes dispatched by type: LLM · condition · passthrough
Runs async; entry = nodes with no inbound edge

LLM layer

LangChain provider interface: ask(prompt, files)
OpenAI · Anthropic Claude · Google Gemini today
Model chosen per node, overridable

API & real-time

FastAPI + Uvicorn (ASGI)
JWT auth · bcrypt · HttpOnly refresh cookie
WebSocket pub/sub streams every step live

Frontend

React 18 + TypeScript + Vite 5
@xyflow/react graph canvas · Zustand state
i18next · TanStack Table · Sass

Behind the scenes

Bring your own model

BYO-model philosophy — no lock-in. Choose models by capability, cost or compliance.

One LangChain-backed interface behind every provider
Available now: OpenAI · Anthropic Claude · Google Gemini
Roadmap: almost any LLM via LangChain — Azure OpenAI, Mistral, Cohere, local / Ollama, and more
Register endpoint + API key; override per node; swap anytime

OpenAIClaudeGemini + more via LangChain

Why it matters

Use the model you trust — and change your mind without rebuilding a single flow.

Your keys. Your data. Your choice.

Behind the scenes

Deploy your way

From a laptop to the cloud — the same app, your choice of footprint.

Desktop

Electron 41 packaged app
Bundles the Python backend
Runs fully local / offline
One-click installer

Self-hosted

Docker Compose, full stack
App + MongoDB in containers
FastAPI serves the built UI
Single process in production

Cloud

Google Cloud Run
Firestore persistence
Autoscaling + HTTPS
Health-probed cold starts

Switchable persistence: JSON · SQLite · MongoDB · Firestore — set by one environment variable.

Roadmap · 2026

What's coming next

Built on pluggable foundations — LangChain providers and swappable backends make each of these an extension, not a rewrite.

More authentication providers

Microsoft Entra ID (Azure AD)
Enterprise SSO via OIDC / SAML

More persistence layers

Azure Cosmos DB
PostgreSQL · Microsoft SQL Server

More input formats

Word (.docx) & Excel (.xlsx)
OpenOffice / ODF (.odt · .ods)

New activity types

Conditional branching
Python expressions · variables

Scheduling & export

Scheduled runs — cloud & self-hosted
Export activity to files / folders & SMB

External REST API calls

Call external services mid-flow
Push & pull data over HTTP