Skip to main content

AI Engine Technical Specifications

Sentra v3 introduces a robust, multi-tenant AI processing pipeline designed to handle proprietary enterprise data with strict access controls and high accuracy. The following specifications detail the core mechanisms powering the Sentra AI ecosystem.

Dynamic RAG & Knowledge Ingestion​

Sentra utilizes a highly configurable Retrieval-Augmented Generation (RAG) architecture tailored for enterprise scale.

1. Vectorized Knowledge Base​

  • Storage: Knowledge embeddings are stored securely utilizing Google Cloud Storage (GCS) buckets, heavily segregated by client_id (Tenant).
  • Crawling & Ingestion: Built-in crawlers process unstructured web data (targeting specific CSS selectors like div.content#main) and local files.
  • LLM Pre-processing: During ingestion, a "Flash LLM" can be configured to pre-process and normalize extracted text using custom prompts before it is vectorized, ensuring high-quality semantic embeddings.

2. Multi-Tenant API Architecture​

  • Single FastAPI Engine: A unified backend (Sentra3/api) processes all incoming LLM queries.
  • Tenant Isolation: The API strictly segregates knowledge bases based on the client_id (e.g., sentra, biggrit). Each tenant has its own isolated Neo4j database prefix (e.g., base_) and GCS bucket ensuring zero cross-contamination of AI context.

Access Controlled Knowledge (RBAC)​

Sentra's AI does not return uniform answers; it is contextually aware of the user's authorization level.

  • OIDC Synchronization: Upon a user's first login via the Sentra OIDC provider, their specific cryptographic roles (ACLs) are synced to the local Sentra database.
  • Role-Based Retrieval: When a query is submitted (e.g., via the Slack integration), the API Gateway first validates the user's role. The Vector Database subsequently filters the embedding search exclusively to documents matching that role (e.g., hiding sentra:admin documents from standard users).

Human-in-the-Loop Validation​

To mitigate the risk of LLM hallucinations in critical business environments, Sentra incorporates a Human-in-the-Loop (HITL) fallback mechanism.

  1. Confidence Thresholds: If the primary LLM cannot retrieve highly relevant chunks from the VectorDB, it flags the response.
  2. Expert Routing: Unknown or flagged queries are automatically routed to a designated human expert within the organization's administration panel.
  3. Continuous Learning: Once the human expert provides the correct answer, that verified response is permanently injected back into the VectorDB. Subsequent identical queries will be handled autonomously by the AI with absolute accuracy.