A full-stack platform for document ingestion, multimodal extraction, prompt-driven workflows, and in-app evaluation.
Built with a React frontend, FastAPI backend, Better Auth sidecar, PostgreSQL, and a flexible AI model layer that can be adapted to multiple providers and deployment environments.
This platform is designed for teams that need to ingest documents, extract structured information, analyze figures, run prompt-based workflows, and evaluate output quality in one place.
It supports:
- Containerized local development and deployment across frontend, backend, auth, and supporting services
- Document processing pipelines with parser options such as Azure Document Intelligence and Docling
- Pluggable model providers including:
- Azure OpenAI
- Google Vertex AI / Gemini
- Ollama
- vLLM-backed endpoints
- other provider-specific or self-hosted runtimes
- Prompt-based entity extraction and structured downstream workflows
- Template workspaces for creating, organizing, versioning, and sharing prompts
- User groups and shared workspaces for collaborative review and reuse
- Improved authentication through Better Auth with GitHub OAuth and Microsoft Entra support
- In-app evaluation workflows powered by DeepEval, including custom evaluation steps and LLM-as-a-judge patterns
- Batch and interactive workflows for extraction, review, and evaluation
- Short-term chatbot memory for follow-up questions within an independent chat conversation
- Production-oriented deployment paths for Azure infrastructure and other containerized environments
Architecture overview. The platform connects document ingestion, the React user interface, FastAPI orchestration, parser outputs, LLM workflows, evaluation, collaboration, persistence, authentication, and observability.
Detailed backend diagrams: Backend visual workflow map.
The application is organized around a small set of core services:
| Service | Port | Purpose |
|---|---|---|
| Frontend (React + Vite) | 3000 | Main user interface |
| Backend (FastAPI) | 8001 | Document processing, extraction, evaluation, and APIs |
| Auth sidecar (Better Auth) | 3001 | Authentication, session handling, and OAuth flows |
| PostgreSQL | 5432 | Application data, auth tables, sessions, templates, and groups |
| Azurite (local dev) | 10000 | Local Azure Blob Storage emulator |
In production, the frontend can be deployed separately while the backend and auth service run together in a containerized environment.
The screenshots below are sourced from
docs/images/.
Figure 1. High-level application flow showing document ingestion, extraction, prompt-driven workflows, review, and in-app evaluation.
Figure 2. Document upload workflow showing support for multiple document ingestion and parser selection through the document parser dropdown.
Figure 3. Expanded figure extraction view with detailed figure analysis and summarization generated by vision large language models (VLLMs).
Figure 4. Entity extraction workflow showing model selection for the extraction step.
Figure 5. Entity extraction page showing the entity name, prompt used for extraction, extracted value, and visual grounding with a bounding box drawn around the cited region in the PDF.
Figure 6. In-app evaluation page where users can enter ground truth and use LLM-as-a-judge workflows to score multiple outputs.
Figure 7. Evaluation results page with a tabular view of scoring outputs and export options for Excel-compatible batch output.
Figure 8. Additional in-app metrics view showing session-level tracking and performance insights captured during workflow execution.
The stack is designed to run in containers for both local development and deployment. The repository already includes Docker support for:
- frontend
- backend
- auth service
- PostgreSQL
- local blob storage emulation
This makes it easier to maintain consistent environments across local development, CI/CD, and cloud deployments.
The application is built so model access is not tied to a single vendor.
The platform is designed to support providers and runtimes such as:
- Azure OpenAI
- Vertex AI / Gemini
- Anthropic-style providers
- Local or self-hosted runtimes such as Ollama and vLLM
This makes it easier to swap providers based on cost, privacy, performance, or deployment constraints.
Authentication is handled by a dedicated Better Auth sidecar backed by PostgreSQL.
This provides:
- GitHub OAuth support
- Microsoft Entra support
- server-side session validation
- clearer separation between application data and auth concerns
- a more portable architecture for self-managed deployments
The platform includes collaboration features for teams:
- create and manage user groups
- share sessions with groups
- browse shared history
- open shared work in a safe copy-on-write workflow
Prompt engineering is a first-class part of the product.
The template workspace supports:
- creating and editing prompt templates
- organizing templates into folders and workspaces
- user, group, and global scope
- sharing templates with collaborators
- template version history and iteration
The app includes built-in evaluation workflows so teams can assess extraction quality without leaving the product.
This includes:
- DeepEval-powered evaluation flows
- configurable metrics
- custom evaluation steps
- support for multiple judge models and providers
- persistent evaluation records tied to sessions
- Docker
- Docker Compose
- Provider credentials for any external models or services you want to enable
From the repository root:
docker compose up --buildThis brings up the local development stack defined in docker-compose.yml, including:
- PostgreSQL
- Better Auth service
- FastAPI backend
- React frontend
- Azurite for local blob storage emulation
- Frontend:
http://localhost:3000 - Backend API:
http://localhost:8001 - Auth service:
http://localhost:3001 - PostgreSQL:
localhost:5432 - Azurite blob endpoint:
http://localhost:10000
The exact credentials and environment variables you need depend on which providers and features you enable.
Common configuration areas include:
- PostgreSQL connection settings
- Better Auth secrets and OAuth configuration
- Azure storage configuration
- Azure OpenAI credentials
- Azure Document Intelligence credentials
- Vertex AI / Gemini project configuration
- optional local or self-hosted inference endpoints
For service-specific setup details, see the linked documentation below.
SummarizationTool-dev/
├── auth-service/ # Better Auth sidecar
├── backend/ # FastAPI API, processing, evaluation, data models
├── frontend/ # React/Vite client application
├── docs/ # Migration, auth, deployment, and internal docs
├── infra/ # Infrastructure and deployment manifests/scripts
├── scripts/ # Deployment and utility scripts
├── docker-compose.yml # Local multi-service stack
└── README.md # Project overview
- Documentation index — start here — full navigation map for all backend, frontend, and deployment docs
- Frontend technical design docs — frontend architecture, page-by-page docs, component index, hooks, and TypeScript interfaces
- Glossary — definitions for all Azure services, tools, and project-specific terms
- Backend README — backend setup and processing details
- Chat memory - chatbot memory behavior, API contract, and operational notes
- Backend technical design docs — backend architecture, workflows, diagrams, data models, schemas, and appendices
- Backend class reference — field-level reference for backend ORM models, schemas, dataclasses, service attributes, and provider classes
- Migration guide — architecture migration and platform transition notes
- GitHub auth setup — Better Auth GitHub OAuth configuration
- Dockerize & deploy to Azure — deployment architecture and implementation notes
This repository is set up for container-based deployment and includes infrastructure artifacts for Azure-based environments.
The current deployment path supports:
- containerized backend and auth sidecar
- separate frontend deployment
- PostgreSQL-backed persistence
- blob storage for uploaded and processed files
- CI/CD-driven image build and rollout
Special thanks to:
- Health Canada Solutions Fund 💖
- Shared Services Canada Science Cloud ☁️