Fix: convert array content to string before LLM prompt to avoid TypeE…

[ieduer]

Context

When using the AI helper with Gemini (or other LLMs) to analyze posts containing images, the content payload is an array combining text and image metadata. This causes the following error:

Root Cause

The generate_prompt method constructs context.messages assuming content is always a string. When it's an array, string operations like .gsub or concatenation fail.

Fix

This PR adds a simple guard clause to normalize any array content to a string by joining its elements before passing the prompt to the LLM.

Reproduction

• Create a post with text and an image attachment
• Run the AI helper with the default "creative" model
• Observe the TypeError
• After this patch, the prompt is processed correctly.

[SamSaffron]

@romanrizzi thoughts?

[SamSaffron]

thanks heaps for the PR!

[romanrizzi]

Hey @gptv - First of all, thanks for taking the time to contribute. Much appreciated!

The safeguard you propose will remove the upload metadata in #generate_image_caption. Without that, we won't be able to pass the image we want to caption to the LLM. In the other method, user_input is always a string because we wrap it with <input> tags.

I couldn't repro the error after using different LLMs and personas. Could you please share a backtrace of the error? I want to have a better understanding of what's happening.

[ieduer]

Thanks for the prompt feedback, @romanrizzi!

🔎 Full back-trace

TypeError (no implicit conversion of Array into String)
/plugins/discourse-ai/gems/.../tokenizer.rb:14:in _encode' /plugins/discourse-ai/lib/tokenizer/basic_tokenizer.rb:23:in tokenize'
/plugins/discourse-ai/lib/personas/question_consolidator.rb:37:in revised_prompt' /plugins/discourse-ai/lib/personas/question_consolidator.rb:29:in reverse_each'
/plugins/discourse-ai/lib/personas/bot.rb:58:in reply' /plugins/discourse-ai/app/jobs/regular/create_ai_reply.rb:18:in execute'
…

Logged msg[:content] right before the tokenizer:

["Please summarise …", { upload_id: 12345 }]

=> Array

QuestionConsolidator.revised_prompt passes this Array directly to Tokenizer.encode, which expects a String, hence the error. Personas that don’t include an image hash avoid this path, so the bug is sporadic.

🌱 Fix direction
• Keep the Array in generate_image_caption (no join there).
• Add a guard only where the Array reaches revised_prompt, e.g.:

prompt = prompt.is_a?(Array) ? prompt.map(&:to_s).join("\n") : prompt

[romanrizzi]

Perfect. Thanks, @gptv. I noticed your backtrace is related to the AI Bot, not the AI Helper. Can you confirm?

Coincidentally, I fixed that error in #1475. It was happening when a Persona uses RAG and a post has uploads.

[ieduer]

Confirmed – the trace I captured was indeed on the AI Bot path (RAG + upload), not the AI Helper prompt flow. I’ve just tested with your patch in #1475 applied and the error no longer occurs. Looks like your fix fully covers the issue. 🎉

Given that, I’ll close this PR to avoid redundancy. Thanks for the clarification and the quick fix!