Consult

Consult is a machine learning and LLM-powered tool to automate the processing of public consultations.

Important

Incubation Project: This project is an incubation project; as such, we don't recommend using this for critical use cases yet. We are currently in a research stage, trialling the tool for case studies across the Civil Service. If you are a civil servant and wish to take part in our research stage, please register your interest here.

Setting up the application

External dependencies

• PostgreSQL (brew install postgresql)
• redis (brew install redis)
• GraphViz (brew install graphviz), for generating diagrams

Environment variables

Populate .env by copying .env.example and filling in required values.

Python

Ensure you have python 3.12.3, poetry and npm installed, then run poetry install.

Database setup

brew services start postgresql

This will run the postgresql locally.

make dev_environment

This will set up dev and test databases with dummy data. See the definition of that make task for the various steps.

make dev_admin_user

This will set up the admin account to dev environment.

You will have an staff user (i.e. one that can access the admin) created with the username [email protected] and the password admin.

Confirm everything is working with

make check_db

(You can see all the available make commands by running bare make or make help).

Run the application

make serve

Run the tests

make test

The database

Generating dummy data

Only run this in development. Will create a consultation with 10 complete responses in a variety of question formats. This runs as part of make dev_environment, but you can run it more than once.

make dummy_data

Or go to /support/consultations/ and generate a dummy consultation from there.

Database migrations and schema diagram

If you use the make migrate command to run migrations, the diagram below will be regenerated automatically. If you need to generate it outside that process, you can run manage.py generate_erd. (You will need graphviz installed: see pydot docs).

Login

Magic links

You can sign into the application using a magic link, requested via /sign-in.

For convenience, in local dev environments the value of the magic link will be logged along with the rest of the server logs.

The frontend

CSS

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

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.

In production, we prepopulate frontend/CACHE using manage.py compress which will mean that every request is served from the cache.

django-compressor also takes care of fingerprinting and setting cache headers for our CSS.

Fonts and images

The govuk assets are versioned in the npm package. make dev_environment includes a step to copy them to the frontend folder from where runserver can serve them; you can rerun this with make govuk_frontend.

Obtaining outputs programatically

The generate_themes command will accept a JSON file containing a ConsultationWithResponses and emit a JSON file containing a ConsultationWithResponsesAndThemes.

Invoke the command like this, replacing the input file with your JSON.

poetry run python manage.py generate_themes --input=tests/examples/chocolate.json --clean --device cpu

Options available for this command are:

--clean: delete this consultation if it already exists in the database. --llm: which llm to use. Pass fake, bedrock, or ollama/model_name. --embdedding_model: pass the model for SentenceTransformers to use in the BERTopic pipeline. If fake is passed, random topics will be generated. --device: pass the hardware device to run the topic embeddings on. Examples include: cpu, cuda and mps. Will default to cpu when requested is unavailable.

The resulting file will be placed in tmp/outputs and its path will be printed on the console.

If you are using Bedrock you will need to assume the ai-engineer-role in your shell before running this command.

If you are using Ollama, you will have to install the app (e.g. brew install ollama) and have it running ollama serve. You will need to run your models e.g. ollama run mistral.

Schema documentation

The data schema for consultations supplied to the tool is defined in consultation_analyser/consultations/public_schema/public_schema.yaml.

To build the JSON schemas and examples from this file, run make schema_docs.

The json-schema-faker-options.js file configures (JSON Schema Faker)[https://github.com/json-schema-faker] to make the JSON examples.