Contrib.social

Option to Suppress Verbose "Thinking" Outputs in DeepSeek-R1 Model

Open on Drupal.org →
Created on 10 February 2025, 12 days ago

Problem/Motivation:

When utilizing the DeepSeek-R1 model within the Drupal AI assistant, the system often produces extensive "thinking" outputs. These verbose responses can overwhelm users and detract from the assistant's functionality, especially when concise answers are more appropriate.

Steps to Reproduce:

  1. Integrate the DeepSeek-R1 model into the Drupal AI assistant.
  2. Initiate a simple query or greeting (e.g., "Hello, AI assistant!").
  3. Observe the assistant's response, noting the inclusion of extensive "thinking" content.

Proposed Resolution:

Implement a configurable option within the Drupal AI assistant settings to suppress or limit the verbose "thinking" outputs from the DeepSeek-R1 model. This feature would allow administrators to toggle the verbosity of responses, ensuring that the assistant provides concise answers unless more detailed information is explicitly requested by the user.

Remaining Tasks:

  • Develop the functionality to control the verbosity of the DeepSeek-R1 model's outputs.
  • Integrate this feature into the Drupal AI assistant's configuration settings.
  • Test the implementation to ensure it effectively suppresses verbose outputs when enabled.

User Interface Changes:

Add a new setting in the Drupal AI assistant's configuration panel, such as a checkbox or toggle switch, labeled "Suppress verbose 'thinking' outputs."

API Changes:

None anticipated.

Data Model Changes:

None anticipated.

✨ Feature request
Status

Active

Version

1.0

Component

AI Assistants API

Created by

🇺🇸United States sushichris

Live updates comments and jobs are added and updated live.
Sign in to follow issues

Comments & Activities

  • Issue created by @sushichris
  • Comment 11 days ago →
    🇩🇪Germany marcus_johansson

    Which provider are you using for it? The Deepseek provider, the Fireworks provider or the Azure provider?

    The provider would have to take care of this, unless its a common architecture.

  • Comment 11 days ago →
    🇺🇸United States sushichris

    Ah, I am using Ollama, I was going to create a bug issue for the ollama provider but I thought it might be more associated with the AI assistant sub module, the assistant module appears to be responsible for the output format. The only Ollama ai model that behaves this way is the deepseek-r1 model, phi4 llama3.1 llama3.2 don't have the "thinking" output so they behave as expected.

contrib.social Blog FAQ Discussions
Production build 0.71.5 2024