01. Overview

Reasonable is a high-performance AI agent service designed to prioritize transparent reasoning over simple task completion. Built on the modern LlamaIndex Workflows architecture, it moves away from traditional "black-box" LLM wrappers to provide a fully observable, event-driven agentic experience.

Instead of a standard chatbot, Reasonable acts as a Socratic Tutor, using diagnostic tools to analyze problem structures and guiding users through identification, strategy, and execution.

---

02. The "Reasoning-First" Philosophy

Unlike traditional agents that "get an answer" in one jump, Reasonable is engineered to show its work.

• Total Transparency: Every step of the agent's internal logic is streamed to the user.
• Scaffolded Interaction: The agent uses tools as "Diagnostic Engines" to gather facts, which it then uses to challenge and guide the user.
• Educational Intent: It focuses on how to arrive at a conclusion, making it a powerful tool for learning and complex problem-solving.

---

03. Core Architecture: Workflows-Native

Reasonable leverages the latest LlamaIndex event-driven architecture (v0.14+). This allows for robust, multi-step agentic behavior (like the ReAct loop) that is far more flexible and resilient than legacy sequential agents.

Event-Driven Logic

The system uses a specialized Event Schema to communicate between the backend and frontend:

• Pulse: Real-time heartbeat of the engine.
• Reasoning: Dedicated channel for "Thinking" tokens.
• Call/Result: Transparent tool interaction logs.
• Message: The final synthesized response.

---

04. Chain-of-Thought (CoT) Extraction

One of Reasonable's standout features is its ability to handle and visualize internal reasoning.

It is specifically engineered to handle "Thinking" tokens from models like Qwen 3.5. It identifies and streams internal reasoning processes separately from final answers, providing total transparency into how the AI arrived at a conclusion.

---

05. The "Socratic" Philosophy

Following its latest architectural pivot, Reasonable has transitioned from a simple "solver" into a Socratic Tutor.

• The Heuristic Ladder: Instead of blurting out answers, the agent uses diagnostic tools to analyze the "shape" of a problem.
• Fact-Based Challenging: It treats tools (like the Math Diagnostic Engine) as sources of objective truth used to guide the user's own reasoning process.

---

06. Technical Highlights

Local-First Intelligence

Reasonable is locally-optimized, currently leveraging models like Qwen 3.5 (9B) via Ollama. This allows for heavy reasoning tasks without relying on external cloud APIs for the "brain" of the operation.

Real-Time Streaming (SSE)

The backend is built with FastAPI, providing full CORS support and Server-Sent Events (SSE). This protocol allows frontends to render complex agentic thoughts and tool interactions in real-time as they happen.

---

07. Tech Stack

Frontend: React, Vite, Tailwind CSS, Framer Motion, Radix UI
Backend: Python, FastAPI, LlamaIndex Workflows
AI Engine: Ollama (Qwen 3.5), ChromaDB
Protocols: SSE (Server-Sent Events)