Skip to main content

AI Engine Architecture

The Bitgrit AI engine represents the cognitive core of Sentra, turning static enterprise data into an interactive, highly secure intelligence layer.

Large Language Models (LLM)​

Dynamic Knowledge Base RAG​

Sentra utilizes an advanced Retrieval-Augmented Generation (RAG) pipeline. Instead of relying on a model's generalized training data, the RAG system connects the LLM directly to your proprietary enterprise data (PDFs, Word docs, internal wikis).

Human-in-the-Loop RAG (Unanswered Queue)​

To ensure maximum accuracy for critical enterprise operations, Sentra supports a rigorous human validation workflow. Generic AI chatbots often hallucinate when they lack information. Sentra prevents this by securely routing unknown queries to human experts.

The Workflow:

  1. Detection & Preservation: If the AI determines it cannot answer a question based on the RAG context (detecting keywords like "I don't know" or calculating low confidence), the query is safely preserved in the database as an UnansweredQuestion.
  2. Pending Queue: The preserved query is routed to the "Unanswered Queue" in the Administrator Inbox.
  3. Expert Feedback: Human instructors review the pending queries and manually draft accurate, verified responses. The admin selects which specific authorization Channel this new knowledge belongs to.
  4. Omnichannel Broadcast: Once resolved, the system immediately pushes the delayed reply back to the original question owner through their connected IM app (e.g., Slack or LINE).
  5. Stored as RAG Knowledge: The verified answer is permanently embedded back into the Vector Database and tied to the selected channel. The next time any user asks the same question, the AI will autonomously retrieve and provide the correct answer.

Pluggable Knowledge Base​

Currently in development, this feature will allow enterprises to plug Sentra directly into self-hosted storage devices, allowing the AI to process data that physically never leaves the client's local intranet.

Public Knowledge Acquisition​

A built-in WebCrawler and screening mechanism allows Sentra to safely ingest and vectorize data from targeted public URLs, combining internal IP with external market data safely.

Access & Integrations​

Role-Based Data Sources (RBAC)​

AI responses are filtered at the database level based on the user's OIDC credentials.

  • A nurse asking a clinical question receives different data context than an HR manager asking an administrative question.

Channel-Based Access & IM Integration​

Users do not need to log into a new portal. Sentra acts as a silent assistant within existing workflows:

  • Available Integrations: Slack, LINE.
  • TBC Integrations: Discord, WhatsApp, Matrix.

Custom AI (Bitgrit Network)​

For tasks that generic LLMs cannot solve, Bitgrit provides custom model development.

Classifiers & Predictive Models​

Utilizing frameworks like TensorFlow, Bitgrit can deploy custom classifiers (e.g., image recognition, complex anomaly detection, demand forecasting) and integrate them seamlessly into the Sentra workflow pipeline.

Through Bitgrit's global community of over 41,000 data scientists, clients can host competitions to source the most accurate, bespoke AI models for highly specialized use cases.