🐳 feat: Deepseek Reasoning UI

[danny-avila]

Summary

Adds simple reasoning UI for LLMs that clearly distinguish reasoning tokens from standard generation tokens (only deepseek-reasoner for now)

Once migrated to new agent framework, will include more metadata such as time elapsed thinking, dynamic thinking status, etc.

Change Type

• New feature (non-breaking change which adds functionality)

Checklist

• My code adheres to this project's style guidelines
• I have performed a self-review of my own code
• I have commented in any complex areas of my code
• I have made pertinent documentation changes
• My changes do not introduce new warnings
• I have written tests demonstrating that my changes are effective or that my feature works
• Local unit tests pass with my changes
• Any changes dependent on mine have been merged and published in downstream modules.
• A pull request for updating the documentation has been submitted.

[christiaangoossens]

Is this compatible with the deepseek-r1 ollama output as well? It uses <think> </think> as markers (even if that's not standard yet).

[michnovka]

Do we need to enable sth special to see the collapsible UI?

[danny-avila]

looks like your frontend is cached, you need to clear cache/localStorage

[danny-avila]

Is this compatible with the deepseek-r1 ollama output as well? It uses <think> </think> as markers (even if that's not standard yet).

no, didn't know it did this. I can add that soon

[danny-avila]

@christiaangoossens Which model/tag are you using with Ollama?

[peterhoang]

looks like your frontend is cached, you need to clear cache/localStorage

I'm getting the same issue. I've tried clearing the cache/localStorage in my Chrome browser (Version 131.0.6778.265).

In the meantime, is there a flag or configuration we can set to force the UI to use this new feature?

[danny-avila]

Hi @peterhoang I see the issue now, I'll be patching it shortly

[christiaangoossens]

@christiaangoossens Which model/tag are you using with Ollama?

https://ollama.com/library/deepseek-r1, afaik all of them use <think> syntax.

[danny-avila]

Got it the mentioned issues are now resolved here: #5456

[xhejtman]

we are almost there with ollama deepseek-r1 (via preset)

[danny-avila]

@xhejtman I recommend using a different title model, since you'll get thoughts in those title generations. I can maybe think of a way to remove them, though

[xhejtman]

@xhejtman I recommend using a different title model, since you'll get thoughts in those title generations. I can maybe think of a way to remove them, though

thanks. However, as of the latest commit, DS R1 with ollama (no openrouter), it does not display thoughts in the main chat window at all, only in sidebar as you noticed.

[danny-avila]

@xhejtman it could be no thoughts were generated, I've noticed this is the case when you just say "hi."

Either this or you need to update Ollama (I thought the same thing until I got the latest version)

[xhejtman]

yes, you are right sorry, it is ok, I was confused by tag in the left sidebar supposing that it actualy returned as it is displayed.

[matthewknill]

All the other models seem to work fine for me except for the deepseek-reasoner model:

Here are the relevant logs:

2025-01-28T00:10:12.183Z debug: [AskController]
{
  text: "What is 500 * 655",
  conversationId: "CONVERSATION_ID_1",
  endpoint: "Deepseek",
  endpointType: "custom",
  resendFiles: true,
  modelOptions.model: "deepseek-reasoner",
  modelsConfig: "exists",
}
2025-01-28T00:10:12.186Z debug: [BaseClient] Loading history:
{
  conversationId: "CONVERSATION_ID_1",
  parentMessageId: "PARENT_MESSAGE_ID_1",
}
2025-01-28T00:10:12.190Z debug: [BaseClient] Context Count (1/2)
{
  remainingContextTokens: 62986,
  maxContextTokens: 63000,
}
2025-01-28T00:10:12.190Z debug: [BaseClient] Context Count (2/2)
{
  remainingContextTokens: 62986,
  maxContextTokens: 63000,
}
2025-01-28T00:10:12.190Z debug: [BaseClient] tokenCountMap:
{
  MESSAGE_ID_1: 11,
}
2025-01-28T00:10:12.190Z debug: [BaseClient]
{
  promptTokens: 14,
  remainingContextTokens: 62986,
  payloadSize: 1,
  maxContextTokens: 63000,
}
2025-01-28T00:10:12.191Z debug: [BaseClient] tokenCountMap
{
  MESSAGE_ID_1: 11,
  instructions: undefined,
}
2025-01-28T00:10:12.191Z debug: [BaseClient] userMessage
{
  messageId: "MESSAGE_ID_1",
  parentMessageId: "PARENT_MESSAGE_ID_1",
  conversationId: "CONVERSATION_ID_1",
  sender: "User",
  text: "What is 500 * 655",
  isCreatedByUser: true,
  tokenCount: 11,
}
2025-01-28T00:10:12.191Z debug: [OpenAIClient] chatCompletion
{
  baseURL: "https://api.deepseek.com/v1",
  modelOptions.model: "deepseek-reasoner",
  modelOptions.user: "USER_ID_1",
  modelOptions.stream: true,
  // 1 message(s)
  modelOptions.messages: [{"role":"user","content":"What is 500 * 655"}],
}
2025-01-28T00:10:12.194Z debug: Making request to https://api.deepseek.com/v1/chat/completions
2025-01-28T00:10:12.200Z debug: [saveConvo] api/app/clients/BaseClient.js - saveMessageToDatabase #saveConvo
2025-01-28T00:11:12.223Z debug: [AskController] Request closed
2025-01-28T00:11:12.224Z debug: [AskController] Request aborted on close

[armikhael]

@michnovka Hi, it gives me the same error. It's not creating the Thinking component. Do I need to do something? I have a fork of the project and pulled the change, but it's not rendering for me.

[danny-avila]

@armikhael looking at your screenshot, it’s clear you don’t have the latest commits

[walbercardoso]

Reference in new is

Same here...

@danny-avila does it work with any model? I tried with o1 and o1-mini... but it's not rendering...

[ErykCh]

o1 and o1-mini they don't return their reasoning

[ErykCh]

The functionality works great for me.

The only thing that could be changed is to automatically close the reasoning section once you start writing the actual answer.

[nikolaypavlov]

Hey @danny-avila Great feature. Thank you. It looks much better now. For whatever reason this feature doesn't work for the "Chat with Files" flow. Is this possible to fix it?