Thinking & Reasoning

Modern AI models use internal "chain-of-thought" reasoning to decompose complex problems, analyze edge cases, and verify hypotheses before formulating their final response.

AIsh.li surfaces this reasoning process in real time through an elegant, collapsible Thought Process interface, providing complete transparency and auditability without cluttering the chat experience.

What is Model Thinking?

When tackling challenging tasks—such as writing complex code, solving multi-step mathematical puzzles, or evaluating multi-condition business policies—advanced models (such as Gemini 3.1 Pro) perform intermediate deductions.

Instead of hiding these steps behind a black box or outputting pages of raw internal monologue, reasoning models generate a dedicated reasoning block:

[thinking]
1. The user asks whether Novello is closer to Barolo or Alba.
2. Looking up coordinates: Novello (44.58° N, 7.92° E), Barolo (44.61° N, 7.94° E).
3. Distance Novello to Barolo: ~3.5 km.
4. Distance Novello to Alba: ~15 km.
5. Conclusion: Barolo is significantly closer.
[/thinking]

Barolo is much closer to Novello, located just 3.5 km away (about a 6-minute drive), whereas Alba is approximately 15 km away.

Real-Time Streaming & Interactive UI

AIsh.li's streaming protocol detects thought tokens (such as [thinking] and <think>) on the fly as they arrive over Server-Sent Events (SSE):

  1. Active Streaming: While the model is actively thinking, the thought drawer automatically expands with a subtle glowing pulse, allowing you to follow its analytical reasoning as it happens.
  2. Auto-Collapse: Once the model completes its thought process and begins generating the final conversational response, the thought drawer cleanly collapses.
  3. Expand on Demand: Users and developers can click the Thought process button at any time to inspect the full chain of thought.
  4. Copy Markdown: Clicking the Copy button on a message copies the clean answer, with an option to inspect raw thoughts.

Why Transparent Thinking Matters

  • Auditability: Verify why an agent made a particular recommendation or chose a specific tool.
  • Error Detection: Spot reasoning oversights or flawed assumptions early in the reasoning chain before taking real-world actions.
  • Clean User Experience: End-users receive a concise, polished response without wading through intermediate scratchpad notes.
  • Debugging Agent Prompts: Observe how the model interprets system instructions and constraints to refine prompts faster.

Try It in Practice

  1. Open Chat with Google Gemini.
  2. Submit a reasoning riddle or puzzle, for example:
    • Math & Trick Logic:

      "A bat and ball cost

      .10 in total. The bat costs
      .00 more than the ball. How much does the ball cost?"

    • Constraint Satisfaction:

      "A farmer has a wolf, a goat, and a cabbage. He needs to cross a river in a boat that can only carry him and one item at a time. If left alone, the wolf eats the goat, and the goat eats the cabbage. How can he cross safely?"

  3. Watch the Thought process accordion expand in real time as the agent works through the intermediate deduction step-by-step!