Agents Overview

NFYio Agents are AI-powered services that process, analyze, and retrieve data. They run within your infrastructure and integrate with your S3-compatible storage, embeddings, and external APIs. Use them to build RAG chatbots, automate workflows, or extend with custom logic.

What are NFYio Agents?

Agents are autonomous services that:

• Process — Ingest documents, chunk them, and generate embeddings
• Analyze — Run LLM inference for summarization, classification, and translation
• Retrieve — Perform semantic search over your document corpus
• Orchestrate — Chain multiple steps with conditional logic and tool use

All agents run in your NFYio deployment, with data staying within your VPC and under your control.

Agent Types

Type	Purpose	Use Case
RAG	Retrieval-Augmented Generation	Document Q&A, knowledge-base chatbots
LLM	Direct LLM interaction	Translation, summarization, classification
Workflow	Multi-step pipelines with tools	Complex automation, agentic workflows
Custom	Your own logic and integrations	Custom tooling, external API integration

RAG Agents

RAG agents combine semantic search with LLM generation. Documents are chunked, embedded, and stored in a vector database. When a user queries, relevant chunks are retrieved and passed to the LLM as context. Ideal for document Q&A, internal knowledge bases, and support chatbots.

LLM Agents

LLM agents interact directly with language models without retrieval. You configure system prompts, temperature, and token limits. Use them for translation, summarization, sentiment analysis, or any task that doesn’t require document lookup.

Workflow Agents

Workflow agents run multi-step pipelines with tool use. They can chain multiple agents, branch on conditions, and call external tools (document search, bucket operations, web fetch). The policy gateway controls which tools each agent can use.

Custom Agents

Custom agents let you build your own agent logic, register custom tools, and integrate with external APIs. Deploy them in VPC private subnets for secure access to internal services.

Architecture

NFYio agents are built on a unified architecture:

┌─────────────────────────────────────────────────────────────────┐
│                     NFYio Agent Service                         │
│                                                                 │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────────┐ │
│  │   RAG       │  │   LLM       │  │   Workflow / Custom     │ │
│  │   Agents    │  │   Agents    │  │   Agents                │ │
│  └──────┬──────┘  └──────┬──────┘  └────────────┬──────────────┘ │
│         │                │                      │                │
│         └────────────────┼──────────────────────┘                │
│                          ▼                                       │
│  ┌─────────────────────────────────────────────────────────────┐ │
│  │              LangChain Integration Layer                    │ │
│  │  • Document loaders  • Embeddings  • Vector stores  • LLMs  │ │
│  └─────────────────────────────────────────────────────────────┘ │
│                          │                                       │
│         ┌────────────────┼────────────────┐                      │
│         ▼                ▼                ▼                      │
│  ┌────────────┐  ┌────────────┐  ┌────────────────────────────┐ │
│  │  GPT-4o    │  │  pgvector  │  │  Policy Gateway & Audit     │ │
│  │  Claude 3  │  │  (embeds)  │  │  • Tool allow/deny         │ │
│  │  Voyage AI │  │            │  │  • Per-step audit trails    │ │
│  └────────────┘  └────────────┘  └────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘

LangChain Integration

NFYio agents use LangChain for orchestration:

• Document loaders — Load from S3 buckets, local paths, or URLs
• Embeddings — OpenAI text-embedding-3-small/large, Voyage AI voyage-3.5-lite
• Vector stores — pgvector for similarity search
• LLMs — GPT-4o, Claude 3, GPT-3.5-turbo via OpenAI-compatible APIs

GPT-4o and Model Support

Default LLM is GPT-4o for high-quality generation. You can configure alternative models per agent. All models are accessed via OpenAI-compatible endpoints, so you can plug in compatible providers.

Policy Gateway

The policy gateway controls which tools and operations each agent can perform:

• Allow/deny per tool — Restrict document search, bucket operations, web fetch
• Per-workspace policies — Different teams get different capabilities
• Audit logging — Every tool call is logged for compliance

Audit Trails

Every agent run produces an audit trail:

• Input query and parameters
• Retrieved documents (for RAG)
• Tool calls and responses (for workflows)
• LLM prompts and completions
• Timestamps and user/workspace IDs

Use audit trails for debugging, compliance, and usage analytics.

Next Steps

• RAG Agents — Build document Q&A and knowledge-base chatbots
• LLM Agents — Direct LLM interaction for translation and summarization
• Workflow Agents — Multi-step pipelines with tool use
• Custom Agents — Build and deploy custom agent logic
• Embeddings & Vector Search — How embeddings and indexing work
• Agent Tools — Built-in and custom tools