Aegis Platform Overview

The Aegis stack is an enterprise-grade framework for building, deploying, and managing LLM-powered agents and retrieval systems. It is designed to help engineering teams move from fast prototyping to robust, secure, and observable AI workflows.

This document provides a high-level tour of the key components in the Aegis architecture and how they interact in production environments.

🔧 Core Design Principles

• Modular microservices: Each component is independent and replaceable.
• Structured agent interfaces: Agents use typed inputs/outputs and can call tools.
• Access-aware RAG: All retrieval honors document-level access control.
• Evaluation-native: Every agent run can be scored, compared, and traced.
• Orchestration-first: Data and agent workflows are both centrally orchestrated.

📦 Platform Components (by Architecture Layer)

1. API Gateway (API GW)

The entry point for all external clients.

• Authenticates requests (Auth0 or API Key)
• Injects tenant/user context
• Forwards calls to agent-core or other internal services
• Enforces RBAC and ABAC

2. Agent-Core

The runtime service for agent execution.

• Loads agent configs and structured schemas
• Runs tools, LLM calls, and team flows
• Interfaces with:
  • LLM Gateway for completions
  • RAG service for contextual grounding
  • Evaluation Service for inline scoring
  • Orchestrator for queue-triggered runs

3. LLM Gateway (LLM GW)

Manages access to external LLM providers.

• Supports OpenAI, Anthropic, Azure, etc.
• Tracks token usage, cost, and rate limits
• Supports stub replay and budget enforcement

4. RAG Service

Structured Retrieval-Augmented Generation pipeline.

• Handles chunking, indexing, and search
• Performs hybrid (vector + lexical) retrieval with Vespa
• Applies strict access filters during retrieval
• Semantic RAG: A key enhancement for enterprise grade workflows. Supports semantic tags generted by agents with links/metadata to other documents

5. Evaluation Service

Scoring and feedback pipeline for agent output.

• Real-time and batch evaluation support
• Built-in metrics (F1, accuracy, precision)
• Stores evaluator logs and run history
• Sends feedback to observability or retraining tools

6. Embedding Server

Internal microservice that handles all embedding operations.

• Embeds documents and chunks before indexing
• Sends vectors to Vespa
• Handles multi-model support if needed

7. Prefect Orchestrator

Manages asynchronous data and agent workflows.

a. Data pipeline:

• Pulls from external sources (S3, GCS, EMR)
• Extracts + cleans sections via Unstructured
• Triggers chunking → embedding → indexing

b. Agent pipeline:

• Triggers multi-agent graph flows from queue
• Stores results and run traces
• Enables scheduled or event-driven execution

8. Postgres DB

Transactional system-of-record for:

• Agent and team definitions
• Run logs and evaluation metadata
• Access control metadata
• Document and chunk metadata

9. Vespa DB

High-performance hybrid retrieval system.

• Stores document embeddings and metadata
• Supports vector + keyword + structured filters
• Ranks results based on customizable profiles

🔄 Data and Execution Flow (Overview)

1. A client sends a request to the API GW.
2. The API GW routes to agent-core with scoped context.
3. Agent-core runs tools, calls LLMs via LLM GW, and retrieves context from RAG.
4. If evaluation is enabled, output is scored and logged.
5. Prefect flows handle background ingestion or scheduled agent runs.
6. Vespa and Postgres store document context and metadata.

📈 Observability and Auditing

• TBC