Open

Open

Components overview

Dashboard / Ingestor Windows Server

Dashboard

.NET website for the end user and admin UI

The Dashboard manages the synced content items, permissions, building collections, chat interfaces, and settings at the user and site levels.

Ingestor service

The Ingestor service is a service with connectors to various data sources. It pulls the content and the permissions and sends them to the Gateway embedding queue. 

Firewall Service

This is a service that Classifies and analyses data at rest against policies.

Firewall API

This is a website exposing API to be used in real-time for inspection and classification.

Gateway Linux Server / Containers

Content loader

The loader extracts the text from the content items and cleans unnecessary parts from the content, such as email signatures and disclaimers in emails.

Embedding service

Gets data from the Loader, splits it into chunks, and transforms the chunked content into vectors.

Each file type requires a different loader to extract the text content. We will use the Unstructured library for most file types.

We will endeavor to keep many file types, paragraphs, and sections in a single chunk, with some context before and after for context. 

The embedding vectors will be stored in a vector DB - we’ve elected to use Postgres Vector BB

Embedding AI model

For the private AI BusinessGPT uses GTE Qwen2 (by AliBaba)

This model supports 29 languages, among them Hebrew :

https://ollama.com/library/qwen2:1.5b-instruct

https://huggingface.co/Alibaba-NLP/gte-Qwen2-1.5B-instruct

According to the MTEB benchmarks (link) this model is better than the one used by OpenAI ( ChatGPT) is is most cost-effective in terms of resources needed

Vector DB

The Vector DB stores embedding as vector content and the chunked text with metadata. 

Embedded vector binary content is also stored here. 

Insights Engine

This business Logic processes questions, searches for relevant organizational content, and interacts with the AI models.

The API gets the query and decides the best algorithm to generate the answer.

Local LLM Linux Server / Container

Customers can choose between using OpenAI’s LLM models such as Chat GPT 4o and Meta's Llama3.1. Llama 3.1 is royalty free for commercial use for for applications with less than 700 million monthly users.

Token classification model

An AI model used for specific data classification, such as PII

Bastion Proxy

Network proxy supporting HTTP and Web sockets needed for analysing the AI services traffic