DEV Community

Cover image for Orka-Reasoning
Mak Sò
Mak Sò

Posted on • Edited on

Orka-Reasoning

Redis AI Challenge: Real-Time AI Innovators

This is a submission for the Redis AI Challenge: Real-Time AI Innovators.

What I Built

I built OrKa (Orchestrator Kit Agents), a modular cognitive architecture that executes YAML-defined reasoning flows using AI agents. Think execution graphs for intelligence: not chatbot logic, but dynamic, stateful, traceable cognition.

OrKa uses Redis as its spine—not just for storage, but as a realtime substrate for multi-agent orchestration, memory routing, and trace replay.

Every AI step is deterministic, observable, and replayable because Redis is not an afterthought—it’s the execution backbone.

Demo

Local UI: https://hub.docker.com/r/marcosomma/orka-ui

SDK: [https://github.com/marcosomma/orka]

How I Used Redis 8

OrKa leverages Redis 8 beyond caching in several core ways:

1. Redis Streams

Every agent execution is logged as a JSON object into Redis Streams. This enables:

  • Full replay of execution trace
  • Reproducible reasoning
  • Cross-agent flow introspection
  • Time-stamped event logging for each decision

Example:

XADD orka:trace * agent_id agent_1 input "What's 2+2?" output "4" timestamp 1723... 
Enter fullscreen mode Exit fullscreen mode

2. Redis as a Coordinated Queue

The orchestrator assigns tasks to agents via Redis Lists (acting as a pub-sub compatible queue system). Agents pull from their specific queues, enabling distributed, scalable execution. Redis here acts as the cognitive message bus.

3. Hash Maps for Scoped Memory

Scoped and ephemeral memory is handled using Redis Hashes, letting agents persist intermediate data accessible across the flow:

HSET orka:memory:session123 agent_1_output "42" 
Enter fullscreen mode Exit fullscreen mode

This supports multi-turn, memory-aware reasoning without relying on external databases.

4. Multi-Tenant Isolation via Namespacing

Redis keys are namespaced per session, tenant, and orchestrator instance. This allows OrKa to run multiple cognitive flows in parallel without collisions, and without touching a traditional RDBMS.

5. Planned Use of Redis Gears

Future versions of OrKa will incorporate Redis Gears to run computation (e.g., scoring or summarizing) directly at the data layer—eliminating latency from roundtrip Python execution.


Redis 8 isn’t just storage in OrKa. It is the nervous system.

Top comments (5)

Collapse
 
prema_ananda profile image
Prema Ananda

Demo
Live UI: orka-ui.web.app

Site Not Found

Collapse
 
marcosomma profile image
Mak Sò • Edited

Thanks! Sorry about that. I "fix it". There is no anymore live UI. Or at least not 24/7. But you can download it locally it will direct connect with local orka infrastructure. hub.docker.com/r/marcosomma/orka-ui
Or just follow readme instruction here: github.com/marcosomma/orka-reasoning

Collapse
 
axrisi profile image
Nikoloz Turazashvili (@axrisi)

aren't we supposed to build something after the challenge has started?

Thread Thread
 
marcosomma profile image
Mak Sò

No idea...

Thread Thread
 
marcosomma profile image
Mak Sò • Edited


Message ca be read as you want.... Anyway Orka-reasoning is work in progress and just 3 months old. I mean you do not build a cathedral over the night...