Skip to main content

Data Flow Overview

Status

Owner: Tech lead · Status: draft · Canonical source: training/diagrams/FleziGalaxy-Architectures/

This document maps how user data, conversation context, documents, and runtime events flow through the Flezi Galaxy platform across their full lifecycle.

Five Primary Data Flows​

  1. Chat & Run Execution: User query flows to Backend, triggers AI Runtime with context, calls LLMs/Sandbox tools, and streams events back via Redis Stream and SSE.
  2. Document & Knowledge Pipeline: Documents uploaded to Blob Storage, chunked, embedded, and indexed into Azure AI Search for hybrid vector retrieval.
  3. Workflow & Human-in-the-Loop: Multi-step DAG execution pauses at approval gates, checkpoints state, and resumes upon human review.
  4. Observability & Governance: Spans, metrics, and token usage are sent asynchronously to MLflow Tracking Server and summarized on the Governance Dashboard.
  5. Third-Party & LLM Integration: Enforces zero-data-retention and no-training policies when interacting with enterprise LLM provider endpoints.

Detailed Flow Pages​

Existing Controls​

  • Clear separation between transient streaming events and durable persisted messages.
  • Sensitive credentials stripped before payload transmission.

Evidence (Files & Configuration)​

  • flezi.app.ai.agent/docs/specs/00-architecture/master-architecture.md
  • flezi.app.ai.agent/docs/specs/01-contracts/event-payload-contract.md

Limitations & Roadmap​

  • Client-side encryption option for enterprise documents before ingestion into knowledge bases.