Leverage optional url loading for label categorization prompts

[AlaeMghirbi]

Why?

In order to centralize our prompts, we have the aiprompts project, which also enables prompt evaluation. However, there is currently no mechanism that directly links the prompts in ai-prompts to the tmail backend. As a result, the prompt must be manually copied and pasted each time, which is inefficient .

What?

We need to implement a direct link between the two projects to retrieve and inject common variables(prompt). The goal is to maintain a single, shared prompt that can be used consistently across different projects.

How?

TO BE DONE ....

[chibenwa]

Please propose a sample configuration

Is this mutually exclusive with configured prompt?

[quantranhong1999]

We need to implement a direct link between the two projects to retrieve and inject common variables(prompt). The goal is to maintain a single, shared prompt that can be used consistently across different projects.

My concern (maybe overkill?): we can not control in advance the prompt, but rather load the prompt at runtime, and could be prone to prompt injection (if someone fk up the content of the AI prompt link).

[chibenwa]

I think it can remain optional loaded at start time for simplicity.

Injection imo is a minor concern.

It would align us with over softs from TWP

[AlaeMghirbi]

Please propose a sample configuration

Is this mutually exclusive with configured prompt?

To ensure coherence with the way the Scribe prompt is fetched, I believe we should also retrieve the classification prompt from the Linagora ecosystem.

I propose exposing the label-classification prompt URL via the Linagora ecosystem discovery endpoint (/.well-known/linagora-ecosystem), configured through the linagora-ecosystem.properties file:

labelClassificationPromptUrl=https://files.twake.app/prompts/label-classification/latest.json 

I suggest that if this URL is defined, the prompt specified in listeners.xml should be ignored or rejected to prevent ambiguity. Ultimately, the systemPrompt entry in listeners.xml could be removed, ensuring a single and centralized configuration managed through AI prompts.

[chibenwa]

Scribe has nothing to do with label categorization. Am I missing something?

Can you please reformulate this proposal so that label categorisation leveragesai prompt repo?

[chibenwa]

I am sorry I misread you. I was convinced your proposal was about Scribe and missed the actual proposal.

However,

• We shall not create unneeded links between featuresin the conf. We might eg deploy the listener without JMAP.
• Duplicating info in the conf is OK we can factorize later in the HELM to have 1 aiPromptUrl injected in 2 places
• I want to retain maximum flexibility and be able to set prompts in the configuration

Can you propose changes to AI configuration in order to add optional loading of prompts at boot time from the configured prompt URL?

Can we introduced an interface dedicated to prompt retrieval ?

On a side note I would rather do the reverse: inject the promptUrl defined in ai conf into the linagora ecossytem though it requires a refactoring... Let's wait to pull this trigger.

[AlaeMghirbi]

Configuration proposal:

In listners.xml:

    <listener>
        <class>com.linagora.tmail.listener.rag.LlmMailClassifierListener</class>
        <configuration>
            <!-- Optional: override the default system prompt -->
            <systemPromptUrl>https://raw.githubusercontent.com/linagora/aiprompts/main/label-categorization/latest.json</systemPromptUrl>
            <systemPrompt>Custom LLM classifier instruction</systemPrompt>
           
        </configuration>
    </listener>

or in ai.properties :

labelClassificationPromptUrl=https://files.twake.app/prompts/label-classification/latest.json

[chibenwa]

Which of those options do we prefer ?

Is the URL alone enough? Don't we need a JSON pointer in order to locate our prompt ?

[AlaeMghirbi]

I’d prefer listeners.xml so the prompt configuration stays scoped to the listener .

The URL response will be a bundle like:

{
  "generatedAt": "...",
  "prompts": [
    {
      "name": "classify-email",
      "version": "1.0.0",
      "description": "...",
      "messages": [
        { "role": "system", "content": "..." },
        { "role": "user", "content": "..." }
      ]
    }
  ]
}

For JSON Pointer:

I think we need to know which entry in prompts[] to use. I don’t think we need a JSON Pointer though; we can keep it simple and explicit by adding a selector in the config for example :

systemPromptName: which prompt to pick (match prompts[].name)

<configuration>
  <systemPromptUrl>https://…/label-categorization/latest.json</systemPromptUrl>
  <systemPromptName>classify-email</systemPromptName>

  <!-- optional fallback -->
  <promptName>…</promptName>
</configuration>