Message Stream Classifier

A websocket client that processes messages from a socket and tries to dentable conversations and categorize them as possible calendar events

High level Information flow

Sync Flow [Deprecated in favour of Async client with last_six_message model for disentanglement]

Async Flow

The Async flow uses qwq:32b llm running with ollama for the disentagling the stream of messages coming in before classification.

Models

There are two problem statements for which we have models for,

Message CLassifier

This is a bert model trained over synthetic data generated for classifying a group of messages to have an intent to get on a call or not.

The saynthetic data generation and the model creation code can be found under notebooks folder

Conversation Disentangelent

The process of converting a stream of messages into conversations is called conversation disentanglement

There are two models,

• Rule Based model, this looks at embedding similarity, user mentions and timestamp differences
• Last Six Message based model this model is based on the assumption that you can solve this problem by just looking at the last 6 messages of the stream and classify the new incoming message as part of either of them or the new one

Running With Docker [Sync Flow Only]

• setup .env file with cp .env.example .env and set the WS_SOCK variable, this is the websocket that gets ingested.
• setup docker if you don't have it yet
• run sh build_and_run.sh
  • what it does
    • creates model (this is where the model gets stored) and results (the output will be generated here) folder
    • it checks if you have the model files if not downloads them
    • run docker build
    • run docker run command with the required attached volumes
    • execute the client.py in docker, this will start processing the stream

Running with manual setup / seeting up for development

• setup the .env file with the WS_SOCK variable
• setup uv https://docs.astral.sh/uv/getting-started/installation/
• create model/bert_classifier_v1 and results folder
• download the model files from
  • config.json
  • tensors
  • training_args
• run uv run ingest

running tests

• after setting up uv, you can run uv run pytest