❒ Kairo Core

Core server for Kairo — an NLP-powered note taker.

Features

Refer to this OpenAPI Spec for the complete list of API endpoints.

How to use this?

Important

Install uv before proceedings.

Before you begin, ensure the following prerequisites are met:

• Run make to install dependencies.

• Duplicate .env.example to .env and set up your READ-ONLY Hugging Face token.

1. Download artifacts (Trained models, datasets, etc.)

make dvc-setup

This will redirect you to browser for OAuth. You might encounter unauthorized warning cause I lazy to KYC for GCloud zzZ

2. Start the local server

make server

This usually starts the server at http://localhost:8000

Note

Pair this server with Kairo Web 🕸️ read more...

How to train your dragon these models?

Caution

This is experimental and pretty compute-intensive.

1. Download associated datasets

make data

This will download the Hugging Face datasets based on this config

2. Training the NER model

uv run -m scripts.train_ner \
    --mode new \
    --batch-size 32 \
    --epochs 10 \
    --learning-rate 1e-4 \

• or you can simply opt out all the configs and run with --mode new only

• include --use-wandb to log the training to Weights & Biases, make sure to login with uv run wandb login first

3. Training the summarization model

uv run -m scripts.train.abs_sum

uv run -m scripts.train.ext_sum

---

want to contribute? feel free to drop a PR or raise an issue!