Plex Recommendations

Feed your Plex recently viewed movies into a RAG powered LLM chain to get recommendations based on your watch history.

Connecing to your Plex

You'll need two things:

1. The IP address that your Plex runs on
2. Your Plex key You can find the Plex key in the preferences.xml file in your installation path. I run Plex on a Synology NAS using Package Center, so I found mine in /PlexMediaServer/AppData/Plex Media Server/Preferences.xml.

Provide both of these to the PLEX_ADDRESS and PLEX_TOKEN environment variables.

Migrating Data

On initial boot, the system will detect if your Plex library is stored in the vector database. If it is not, your media will be retreived. You need to provide the default library to download media from via the PLEX_DEFAULT_LIBRARY_SECTION environment variable. You will need to query Plex yourself to get this, but for me, my movies are in section 3. I will fall back to this section if you do not provide one.

Connecting to your LLM

This recommendation engine connects to Ollama. You can bring your own or run it on the cloud. Just provide OLLAMA_ADDRESS, OLLAMA_EMBEDDING_MODEL, and OLLAMA_LANGUAGE_MODEL as environment variables.

Grounding your LLM

You can find the RAG prompt in backend/internal/pkg/langchain/generate.go. This is written to my specific needs. If your needs are not my needs, adjust the grounding prompt accordingly. Note that if you have a very large media collection that it may exceed the context window of the model you are using. You can adjust the amount of titles you retreive by adjusting the limits passed into the media getters in backend/internal/pkg/plex/api.go.

Building and Running

Compiling from source

Download this repository and build the app using go build -o recommendations .backend/cmd/main.go. This builds a binary called recommendations that you can run with ./recommendations.

Cached Responses

A Postgres database is attached at ./pg-data and is used to cache recommendations. The titles provided from recently viewed get base-64 encoded and stored along with the normalized recomendation provided by the LLM. This cache is looked up before requesting LLM recommendation and is returned as is if one is found.

Default Postgres Environment

The default Postgres environment is as follows:

POSTGRES_PASSWORD=postgres
POSTGRES_USER=postgres
POSTGRES_DB=caches
POSTGRES_PORT=5432
POSTGRES_HOST=postgres

You may override this by providing your own values to these environment variables

Docker Compose

There is a docker-compose.yml file in the root of this repo. You can run the Recommendation Engine and the associated Weaviate database by issuing the TAG_VERSION=foo docker compose up --build command. Make sure you have Docker and Docker Compose installed.

Testing this app

There are several tests in the internal packages that were almost all written by an LLM. You can test this program using go test ./... from the root of this repo. These tests are automatically run when you build with Docker.

Open Telemetry

Open Telemetry tracing is instrumented in the backend. To use this out of the box, set OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=otlp://jaeger:4317 in the environment and ensure that the Jaeger service is started from Docker Compose. Traces are collected in Jaeger and accessible at http://localhost:16686. Metrics are available to extend by passing the telemetry.WithMeter option telemetry.InitOtel(). You will have to instrument metrics yourself.

Disabling Telemetry

Telemetry can be disabled by setting DISABLE_TELEMETRY=true in your environment. Note that if using docker compose up, the Jaeger container will still start.