Skip to content

Latest commit

 

History

History
68 lines (47 loc) · 3.59 KB

README.md

File metadata and controls

68 lines (47 loc) · 3.59 KB
Raw

Peak Genie (All features are not available in this repo, just initial version)

Information retrieval on your private data using Embeddings Vector Search and LLMs.

Architecture

image description

Slack bot flow

image description

QA ChatBot

  1. The data is first scraped from the internal sources.
  2. The text data is then converted into embedding vectors using pre-trained models.
  3. These embeddings are stored into a Vector database, we have used Chroma in this project. Vector databases allows us to easily perform Nearest Neighbor Search on the embedding vectors.
  4. Once all data is ingested into the DB, we take the user query, fetch the top k matching documents from the DB and feed them into the LLM of our choice. The LLM then generates a summarized answer using the question and documents. This process is orchestrated by Langchain's RetrievalQA chain

Slack Bot

  1. The package also has an api folder which can be used to integrate it as a Slack slash command.
  2. The API is built using FastAPI, which provides a /slack POST endpoint, that acts as URL for slack command.
  3. Since the slash command has to response within 3 seconds, we offload the querying work to Celery and return processing response to user.
  4. Celery then performs the retrieval and summarization task on the query and sends the final response to the Slack provided endpoint.

Installation

Install using pip

The package can be easily installed by pip, using the following command:

pip install info_gpt[api] git+https://github.com/{}/info_gpt

Development install

  1. This package uses Poetry for dependency management, so install poetry first using instructions here
  2. [Optional] Update poetry config to create virtual environment inside the project only using poetry config virtualenvs.in-project true
  3. Run poetry install --all-extras to install all dependencies.
  4. Install the pre-commit hooks for linting and formatting using pre-commit install

Usage

All configurations are driven through the constants.py and api/constants.py. Most of them have a default value but some need to be provided explicitly, such as secrets and tokens.

Data ingestion

from info_gpt.ingest import Ingest
import asyncio

ingester = Ingest()
# ingest data from GitHub
asyncio.run(ingester.load_github("<org_name_here>", ".md"))
# ingest data from Confluence pages
ingester.load_confluence()

Run API locally

  1. Build the Docker image locally using docker build --build-arg SLACK_TOKEN=$SLACK_TOKEN -t info-gpt .
  2. Run the API using Docker Compose docker compose up
  3. You can use ngrok to expose the localhost URL to internet using ngrok http 8000 --host-header="localhost:8000"

Deployment

This Service definition creates a "ClusterIP" type Service named "peak-genie-service." It will route incoming traffic from port 80 to the Pods running the "api" component of the "peak-genie" application, which are selected based on the labels "app: peak-genie" and "component: api." The Service makes the "api" component accessible within the cluster using the Service's internal IP address. Other applications within the cluster can communicate with the "api" component using the Service's name ("peak-genie-service") and port number (port 80).

Deploy on K8s using Helm chart

---- WIP ----