eru-ai — Agentic AI, RAG & Tool Calling

Why eru-ai

AI apps that act, not just answer.

Wiring agents to real systems is the hard part — tools, schemas, retries, memory, multi-step plans, and a different SDK per model. eru-ai gives you one configurable runtime over Anthropic, OpenAI, Bedrock and Gemini.

Built-in agent types, RAG, conversation memory, and the ability to trigger eru-functions workflows in-process mean you define an agent in JSON, swap models with a config, and ship.

3

agent types

4

LLM providers

20+

tool categories

5

RAG strategies

Core features

Built for agentic production workloads

3

Three agent types

REFLEX, REASONING, ORCHESTRATOR — the right tool per task.

⟲

Reasoning loop with traces

Iterative think → act → observe with extended thinking (Anthropic) and reasoning effort (OpenAI o-series). Every step captured.

⌘

Multi-agent orchestration

An OrchestratorAgent decomposes tasks into sub-agent workflows with shared Blackboard state — parallel or sequential.

{ }

Tool calling, 20+ categories

Banking, KYC, e-commerce, messaging, analytics, e-sign, repos and more — JSON-schema validated, dependency-aware.

≡

Five RAG chunking strategies

Fixed, recursive, sentence, paragraph, semantic — with overlap, structure-preserving separators and language hints.

∞

Conversation memory

Token-aware history with summarization, TTL persistence and per-model budget profiles — production-safe by default.

⌗

Semantic memory across sessions

Plug a vector store — Pinecone, ChromaDB, pgvector, S3 Vectors — and agents recall context across conversations.

↯

Embedded FuncGroups

Agents carry full eru-functions workflows as embedded resources and trigger them in-process — no network, no glue.

~

Streaming & reasoning interfaces

Streaming Model and Reasoning Model extensions on every provider — bring your own model, same contract.

Providers & backends

Plug into your AI stack

Anthropic Claude OpenAI AWS Bedrock Google Gemini Pinecone ChromaDB pgvector S3 Vectors OpenAI Embeddings HuggingFace Bedrock Embeddings

Configure, don't code

An agent in JSON

No SDK lock-in. Swap providers, models, tools or memory by editing config.

agent.json

// REASONING agent with tools + semantic memory
{
  "agent_type": "REASONING",
  "name": "invoice_extractor",
  "model": {
    "provider": "anthropic",
    "model_id": "claude-sonnet-4-6",
    "extended_thinking": true
  },
  "tools": ["ckyc_verify", "processo_sql"],
  "semantic_memory": {
    "vector_store": "pinecone",
    "namespace": "invoices"
  },
  "conversation": { "max_turns": 20, "summarization_threshold": 12 },
  "embedded_funcgroup": "invoice_post_process"
}