Bug | AI Chatbot Assumes Files Have Been Uploaded

[Petar_Dietrich]

Today, I discovered that certain statements or queries entered into the chatbot will trigger it to assume files have been uploaded - plus - in some instances it reveals the file names of the files contained in our configurator’s KNOWLEDGE SOURCES > Files section. See below.

Can you look into this to determine what’s happening? My first guess is that your new Attachment feature code (instructions) are allowing the chatbot to hallucinate.

A quick test is to ask your chatbot (without uploading any files): “Did you get my uploaded files?” or “I uploaded a file. Did you get it?”

Thank you!

---

[Max]

Hi, Petar!

Please let me look into this with the devs. I’ll report back once I have any news

[Petar_Dietrich]

Test on the front end, not the back end (via the configurator). Cheers!

[Max]

Hi @Petar_Dietrich

Thank you for waiting!

The issue you experienced was specific to the current language model (GPT-5 mini). We’re planning to upgrade all users to a more advanced model (GPT-5.4-nano) next week.

In the meantime, we’ve already switched your widget to GPT-5.4-nano to fix the issue. Please give it a try and let us know if the bot is responding correctly now

[Petar_Dietrich]

Hi @Max,

Unfortunately, I cannot test the potential solution on the front end because I am out of tokens. However, based on my back-end testing, the issue remains. I also tested Elfsight’s company chatbot on the front end and received some interesting answers. See below for results.

In my opinion, the chatbot needs further training to handle edge cases like this one. Since this training needs to be incorporated by your developers, I am not sure there is anything your customers can do on their end. If there is, please provide instructions that we can incorporate into the configurator.

Thank you!

---

[Max]

Got it, thanks!

I’ve shared your feedback with the devs and will get back to you once I have their response

[Petar_Dietrich]

Hey @Max,

No rush. Simply wondering if you received any updates from your developers.

Also, please know that this issue – even with GPT-5.4 nano – has not yet been fixed. See below for chatbot response I received today (back end query).

Cheers!

---

[Max]

Hi there, @Petar_Dietrich

Apologies for the wait!

I’ve requested an update from the devs and will let you know as soon as I have any news

[Max]

Hi @Petar_Dietrich

Thank you for waiting!

Our devs have investigated the reported behavior and, unfortunately, we can’t guarantee 100% accuracy for now.

Since the assistant’s answers to the same question may vary significantly, it’s hard to find a single prompt that would completely fix the issue. The issue may be related to the size and specificity of the instructions in your widget. In particular:

The section ### 0.8.2 ACTION-ORIENTED ATTACHMENT TRIAGE (MANDATORY EXECUTION) may be contributing to the issue, as it can sometimes trigger even with the #14. NO SPECULATION & VALIDATION block in place.

Adding a PRECONDITION could help reduce this:

### 0.8.2 ACTION-ORIENTED ATTACHMENT TRIAGE (MANDATORY EXECUTION)

PRECONDITION (HARD GATE — applies before everything else in this section):
This section applies ONLY when a permitted file is ACTUALLY attached to a message in the current conversation (its content is present inline in the chat).
- A user's verbal claim that they uploaded files (e.g., "did you get my files?", "summarize my uploads") is NOT sufficient and MUST NOT trigger triage.
- If no attachment is actually present, the assistant MUST NOT acknowledge, confirm, summarize, triage, or search for any file. It MUST briefly and neutrally ask the user to attach or paste the content, then stop.
- The assistant MUST analyze ONLY the inline content of the actual attachment, and MUST NEVER use the internal knowledge base or file search to locate, confirm, substitute, name, or describe the user's attachments. Internal knowledge MUST NEVER be presented as the user's uploaded files.

When a permitted file (PDF, TXT, MD, CSV, PNG, JPG, WEBP) is uploaded, the assistant MUST NOT provide a passive acknowledgment. It must immediately execute a triage based on the file type:

Also we recommend strengthening the gate in section 14 (last paragraph):

The assistant must not introduce external interaction events (e.g., file uploads, emails sent, documents received, attachments provided) unless the file or event is actually present in the conversation (its content is included inline). A user's statement, request, or claim that they uploaded, sent, or provided something does NOT, by itself, establish that any file or attachment exists.

And the gate in section 15.3.

 15.3 The assistant must never assume user-provided documents, uploads, or attachments exist unless their content is actually present inline in the message. A user's verbal claim that they uploaded something is NOT sufficient.

Our investigation shows that 80% of responses are correct without any prompt changes.

We’ll continue testing the issue on our “clean” version of the assistant and will update you if we find any progress

[Petar_Dietrich]

Thank you, @Max!

I will review, apply, and test the changes. If the issue continues or gets worse, I’ll let you know.

Quite frankly, the detection of uploaded files during the chat session (and respective response) should be part of Elfsight’s AI Chatbot code, not the customer’s. Otherwise, all of your customers are going to have to train their chatbot while unnecessarily burning through message tokens.

Are you guys looking into that?

Again, thank you!

[Max]

Yes, I totally get your point, but since AI isn’t perfect yet, the assistant doesn’t always recognize when files haven’t been uploaded.

We’re exploring if it’s possible to tweak this behavior and will let you know if there are any updates.

Thank you for understanding

[Petar_Dietrich]

Hey, @Max,

I added new instructions (similar to yours) to our AI chatbot configurator. I will test them and report my findings soon after our AI chatbot widget is re-activated (we’re out of tokens).

Cheers!