Issue with AI Auto-Fill Data Overlap
Hi @BuildShip team,
I am using the OpenAI Assistant node with a tool that makes an API POST request to save the name, email, and phone number of the user. The name, email, and phone number are all auto-filled by the AI. While this functionality generally works well for collecting this information during conversations with users, I have encountered an issue.
I have attached a screenshot as an example.
On the left side of the screenshot, I tested this workflow as John Doe. The AI successfully collected and saved the name, email, and phone number during the conversation using the 'Update Contact' tool. Everything seemed to be functioning correctly for this test run.
However, in a subsequent test, I used the workflow as Jane Doe. As shown on the right side of the screenshot, after providing the name Jane, the AI used the 'Update Contact' tool to save the name in our CRM. Unfortunately, it also saved the previous test run's information for John Doe. This should not have occurred, as these were completely separate instances run in different incognito browser windows. The thread IDs for each contact are different, indicating that the AI does not remember previous conversations, but for some reason, the AI auto-fill retained and used the information from John Doe’s session.
This is a significant issue because our CRM automatically merges contacts with the same email or phone number, which means that contacts collected by the AI could be incorrectly merged, even if they originate from different instances. I am concerned about how the AI auto-fill retains information from separate thread IDs.
Could you please advise on how to resolve this issue? It is critical for us to ensure that the system is robust, especially given the high volume of traffic we expect.
P.S. I already sent a support ticket but wanted to share this here for anyone experiencing the same issue.
Solution:Jump to solution
@Jrocks, On checking the recent and the previous logs (which shows the issue) of your shared workflow via support the newer logs shows the correct values and does not repeats, As for the new logs if you compare the request inputs - the request body
email matches with the email of your Update Contact node's input. On latest logs.
We have found an issue that in some of the previous executions it used previous test data and repeated the executions twice, we are looking into this and will add a fix.
For now as a workaround, could you please delete and add the Assistant node again and give a try, and ship it again, it should perhaps solve this issue. Let us know if it works and occurs again....6 Replies
AI Support Bot Information
<@376223473003003904> you can react on the relevant answer (message) with a ✅ in this thread when you think it has been solved, by the bot or by a human!
Anyone can react on this message with a ❌ if the GPT bot is unhelpful or hallucinating answers.
Please note: Team members will review and answer the questions on best efforts basis.
AI Response
This issue seems related to how the OpenAI Assistant node is retaining information across different sessions. From what you've described, it should ideally not carry over any information from one session (or thread) to another.
Here's a possible solution to investigate and correct this issue:
1. Session Management: Ensure that each session is truly isolated. You might need to review the code or configuration managing session IDs and how they are handled throughout the workflow.
2. Resetting State: Explicitly reset the Assistant's state or context between different sessions. This way, any retained information from the previous session is discarded. Look at your workflow configurations to see if there’s an option to clear or reset states.
3. Contact Matching & Deduplication: Implement robust contact matching and deduplication logic within your CRM system. This might help mitigate issues where inaccurate merges happen due to overlapping contact information.
In BuildShip, to isolate context, especially for nodes like the OpenAI Assistant, you need to ensure that each node invocation is stateless or appropriately resets the state. If the node indeed supports such a reset, the configuration should be referred to in the BuildShip documentation for the OpenAI Assistant node.
While detailed troubleshooting might need access to specific code or configurations, consider referring to BuildShip documentation for best practices related to state management and API integrations.
If these suggestions do not solve the problem, or if you need further customization, submitting a detailed support ticket with your workflow details will help the BuildShip support team assist you more accurately.
Make sure you refer to robust session management practices, as mentioned in the BuildShip forums, to avoid cross-session data retention【4:16†source】.
BuildShip AI Support
I've tested this several times, even re-creating the workflows and breaking them down into much smaller versions of the original. I've concluded that the issue is with BuildShip's AI Auto Fill. Somehow, it saves the previous data in its memory and reuses it in the next instance, regardless of threadId.
I even played around with the AI prompt and the AI Auto Fill prompt, but the same issue still occurred. I hope this gets fixed soon, as it can't be used as a production backend for what I'm trying to do, especially considering that a lot of traffic will go through this endpoint once it's in production and released to clients.
Hi @Jrocks thanks for sharing via support, we'll check and update you soon.
Solution
@Jrocks, On checking the recent and the previous logs (which shows the issue) of your shared workflow via support the newer logs shows the correct values and does not repeats, As for the new logs if you compare the request inputs - the request body
email matches with the email of your Update Contact node's input. On latest logs.
We have found an issue that in some of the previous executions it used previous test data and repeated the executions twice, we are looking into this and will add a fix.
For now as a workaround, could you please delete and add the Assistant node again and give a try, and ship it again, it should perhaps solve this issue. Let us know if it works and occurs again.Thanks! The workaround worked.