PostHog-LLM-Lite

PostHog-LLM-Lite is a fork of PostHog with extensions to use PostHog as a LLM text analytics platform and be able to display the LLM interactions while exploring insights.

This lite version is a fork of PostHog v1.31 that only uses Postgres and Redis (current PostHog also uses Clickhouse, Kafka and a few other components). The latest PostHog is undoubtly better and we published LLM extensions for it in PostHog-LLM.

We find this version interesting because of its simplicity, using a straightforward stack of Django, Celery, Postgres, and Redis, making it an good starting point for development and experimentation. It is also simpler and needs less resources for deployment.

Get started with PostHog-LLM-Lite

These are instructions to setup and use this version of PostHog. A good source of additional documentation is this version of the posthog.com documentation repository.

Run on Docker

On a system with Docker and Docker compose installed, first clone the repository, then first generate a random key (eg. run openssl rand -hex 32 to generate a key). Then, open the docker-compose.yml file and replace "<randomly generated secret key>" for the key you generated. This means the SECRET_KEY: "<randomly generated secret key>" line will end up looking something like this (with your key, of course):

```
SECRET_KEY: "cd8a182315defa70d995452d9258908ef502da512f52c20eeaa7951d0bb96e75"
```

Then, to run do:

docker-compose up -d

PostHog should be accessible on the domain you set up or the IP of your instance.

Setup for local development

Clone the repository:

git clone https://github.com/postlang/posthog-llm-lite

Make sure you have Python 3.10 installed python3 --version.

Make sure you have Docker and Docker Compose installed, then run docker compose to start Postgres and Redis.

docker compose -f docker-compose.dev.yml up -d redis db

Navigate into the correct folder (project's root directory):

cd posthog

Create the virtual environment in current directory called 'env':

python3 -m venv env

Activate the virtual environment:

source env/bin/activate

Install requirements with pip

pip install -r requirements.txt

Install dev requirements:

pip install -r requirements-dev.txt

Export the connection to the database

export DATABASE_URL=postgres://posthog:posthog@localhost:5432/posthog_lite

Run migrations

DEBUG=1 python3 manage.py migrate

Prepare the OS with the following dependencies:

1. Install nvm: https://github.com/nvm-sh/nvm

2. Install the latest Node.js 16 with nvm install 16. You can start using it in the current shell with nvm use 16.

3. Install yarn with npm install -g yarn@1.

Install the packages and build the app. Also build the plugin-server.

yarn install
yarn build

cd plugin-server  
yarn install
yarn build

Start the backend, worker, and frontend simultaneously with:

./bin/start

Now open http://localhost:8000 to see the app.

Install Plugins

To maximize the capabilities of PostHog-LLM-Lite for analytics in LLMs, it's important to augment LLM interactions with additional metadata. This enhancement can be done through the deployment of machine learning models as PostHog plugins. Examples of such plugins, which are available in their respective repositories, include:

LLM conversation flow classifier: Enhances understanding of the conversation with specific agent and user flow labels.

Plugin: https://github.com/minuva/ph-flowdetect-plugin

Model: https://github.com/minuva/llm-flow-classification

Emotion classifier: Identifies emotions in text, providing a deeper insight into the emotional context of interactions.

Plugin: https://github.com/minuva/ph-emotion-plugin

Model: https://github.com/minuva/fast-nlp-text-emotion

Toxicity classifier: Detects toxic content in text, allowing for a safer and more positive user experience.

Plugin: https://github.com/minuva/ph-toxicity-plugin

Model: https://github.com/minuva/fast-nlp-text-toxicity

These plugins link the PostHog data ingestion process with the machine learning models. These models need to be hosted on separate servers, a setup process that is detailed within each model's repository.

To install a plugin, navigate to the 'Plugins' section within the PostHog-LLM-Lite user interface, select the 'Advanced' tab, and enter the GitHub repository URL of the plugin into the designated field. This action will fetch the plugin's code from GitHub and initiate its installation. Subsequently, specify the address and port of your model server in the 'HOSTNAME URL' field to complete the setup.

PostHog's plugin system is designed to facilitate the configuration of a model pipeline, enabling sequential data processing where each model can access and use properties appended by preceding models. The order of these models, as displayed in the PostHog UI, is important for ensuring that each model operates on the enriched dataset provided by its predecessors. PostHog-LLM-Lite adheres to a naming convention for LLM properties, and the machine learning models above are triggered exclusively for LLM events that include interaction texts.

Check PostHog

Check the PostHog repository for more details all for the list of authors and contributors.

Open-source

This repo is MIT licensed.