📮 Redbox Copilot

Important

Incubation Project: This project is an incubation project; as such, we DON’T recommend using it in any critical use case. This project is in active development and a work in progress. This project may one day Graduate, in which case this disclaimer will be removed.

Note

The original streamlit-app has moved to its own repository https://github.com/i-dot-ai/redbox-copilot-streamlit.

Redbox Copilot is a retrieval augmented generation (RAG) app that uses GenAI to chat with and summarise civil service documents. It's designed to handle a variety of administrative sources, such as letters, briefings, minutes, and speech transcripts.

Better retrieval. Redbox Copilot increases organisational memory by indexing documents
Faster, accurate summarisation. Redbox Copilot can summarise reports read months ago, supplement them with current work, and produce a first draft that lets civil servants focus on what they do best

intro.mp4

Setup

Please refer to the DEVELOPER_SETUP.md for detailed instructions on setting up the project.

Codespace

For a quick start, you can use GitHub Codespaces to run the project in a cloud-based development environment. Click the button below to open the project in a new Codespace.

Development

You will need to install poppler and tesseract to run the worker

brew install poppler
brew install tesseract
Download and install pre-commit to benefit from pre-commit hooks
- pip install pre-commit
- pre-commit install

Testing

Unit tests and QA run in CI
At this time integration test(s) take 10+ mins to run so are triggered manually in CI
Run make help to see all the available build activities.

Dependencies

This project uses a microservice architecture.

Each microservice runs in its own container defined by a Dockerfile.

For every microservice that we have written in python we define its dependencies using https://python-poetry.org/.

This means that our project is structured approximately like this:

redbox-copilot/
├── frontend/
├── django_app
│  ├── app/
│  ├── static/
│  ├── tests/
│  ├── manage.py
│  └── Dockerfile
├── worker
│  ├── src/
│  │  └── app.py
│  ├── tests/
│  └── Dockerfile
├── redbox/
│  ├── exceptions/
│  ├── export/
│  ├── llm/
│  ├── models/
│  ├── parsing/
│  ├── storage
│  ├── tests/
│  └── Dockerfile
├── docker-compose.yaml
├── pyproject.toml
├── Makefile
└── README.md

Contributing

We welcome contributions to this project. Please see the CONTRIBUTING.md file for more information.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Security

Important

The core-api is the http-gateway to the backend. Currently, this is unsecured, you should only run this on a private network.

However:

We have taken care to ensure that the backend is as stateless as possible, i.e. it only stores text chunks and embeddings. All data is associated with a user, and a user can access their own data.
The only user data stored is the user-uuid, and no chat history is stored.
We are considering making the core-api secure. To this end the user-uuid is passed to the core-api as a JWT. Currently no attempt is made to verify the JWT, but in the future we may do so, e.g. via Cognito or similar

You can generate your JWT using the following snippet. Note that you whilst you can use a more secure key than an empty string this is currently not verified.

from jose import jwt
import requests

my_uuid = "a93a8f40-f261-4f12-869a-2cea3f3f0d71"
token = jwt.encode({"user_uuid": my_uuid}, key="")

requests.get(..., headers={"Authorization": f"Bearer {token}"})

You can find a link to a notebook on how to generate a JWT in the here.

If you discover a security vulnerability within this project, please follow our Security Policy.

Troubleshooting

Error: Elasticsearch 137

ERROR: Elasticsearch exited unexpectedly, with exit code 137

This is caused by Elasticsearch not having enough memory.

Increase total memory available to 8gb.

colima down
colima start --memory 8

Error: Docker... no space left on device

docker: /var/lib/... no space left on device

This is caused by your own laptop being too full to create a new image.

Clear out old docker artefacts:

docker system prune --all --force

Frontend

CSS

We depend on govuk-frontend for GOV.UK Design System styles.

npm install

Once this has been done, django-compressor should work automatically to compile the govuk-frontend SCSS on the first request and any subsequent request after the SCSS has changed. In the meantime it will read from frontend/CACHE, which is .gitignored.

When we get to production, we can prepopulate frontend/CACHE using manage.py compress before building our container, which will mean that every request will be served from the cache.

django-compressor also takes care of fingerprinting and setting cache headers for our CSS so it can be cached.

Fonts and images

The govuk assets are versioned in the npm package. On initial app setup you will need to run poetry run python manage.py collectstatic to copy them to the frontend folder from where runserver can serve them.

We’ll revisit this process when we deploy the app.

How to deploy

checkout the main branch of the following repos:

If, and only if, you want to deploy something other than HEAD then replace var.image_tag in infrastructure/aws/ecs.tf with the hash of the build you want deployed.

Now run the commands below remembering to replace ENVIRONMENT with dev, preprod or prod

cd redbox-copilot
make tf_init
make tf_apply env=<ENVIRONMENT>

dgem/redbox-copilot