yeahhe365/Prisma

What it solves

Prisma is an experimental application designed to explore the limits of multi-model reasoning. It transforms the standard chatbot experience into a "visual thinking laboratory," allowing users to see how complex problems are decomposed and solved by multiple specialized AI agents.

How it works

Prisma uses a "Deep Multi-Agent Reasoning" architecture that follows a specific workflow:

1. Manager Agent: Analyzes the user's query and dynamically plans which specialized "expert" roles are needed.
2. Expert Pool: Generates and executes specialized AI roles (e.g., logicians, architects) in parallel to handle sub-tasks.
3. Recursive Refinement: The Manager reviews the experts' output; if gaps are found, it can trigger a second round of correction.
4. Synthesis: Aggregates the findings from all experts into a final, cohesive response.

Who it’s for

It is designed for users who want to move beyond simple chat interfaces to understand the internal reasoning process of AI, as well as those using Gemini or OpenAI-compatible models to solve complex, multi-step problems.

Highlights

• Full-Chain Visualization: Displays the reasoning process via a dynamic node graph (Analyzing $\rightarrow$ Working $\rightarrow$ Reviewing $\rightarrow$ Synthesizing).
• Transparent Thinking: Users can inspect the "Internal Monologue" and drafts of each individual expert agent.
• Thinking Budget: Allows users to set the depth of thinking (Minimal to High) for the planning, execution, and synthesis phases.
• Local-First Privacy: Stores chat sessions in the browser's IndexedDB and settings in localStorage.
• Broad Model Support: Compatible with Gemini and OpenAI-compatible APIs, including support for images, PDFs, and audio/video (via Google models).