The first phase of the summer vacation program was not displayed, and the mid-week schedule for the fall semester was not mentioned.
Hello, we have received your question. A colleague has been assigned to investigate!
Hello! We have identified the issue.
Cause of the problem:
For the current “Municipal Federation of Trade Unions Campus,” the file “Processed Course Data for 6 Major Campuses.txt” in the knowledge base was split into over 30 chunks. We set the maximum number of chunks to retrieve from the knowledge base to 10, and the 3ChatAgent platform supports a maximum retrieval of 20 chunks. Consequently, not all courses related to the Municipal Federation of Trade Unions were retrieved, leading to some course omissions.
You can resolve this by modifying the knowledge base using one of the following methods:
Method 1:
- Merge the current chunks to reduce the total number of chunks.
- Increase the number of chunks retrieved.
This ensures that all relevant knowledge can be retrieved, thereby enabling a complete answer.
Method 2:
Create a new knowledge base file containing a summary of the total course schedule for each campus. Use one chunk per campus, where each chunk records: “Course schedule for [Campus Name]: [Specific course content].”
Details:
The current knowledge base retrieval settings are shown in the image below. We have set the maximum retrieval to 10 chunks, while the 3ChatAgent platform system supports a maximum retrieval of 20 chunks.
The image below shows the actual retrieval during operation, which retrieved 10 chunks.
I have reorganized the course introduction, keeping the number of slices to under 10. However, there are still errors in the course organization. For example: As shown in the image below, our No. 7 Middle School course is scheduled for 14:00–16:00 in the first session, but the AI response does not reflect this.
Hello, an engineer has been arranged to troubleshoot and debug this morning. We will promptly share the results in the community.
Hello, please check it in your workspace.
Issue 1:
The reason here is that for the image sent by the customer regarding the current issue, we prioritized using the information from the image to answer. Due to the special characteristics of tables in the image, during image understanding, it was interpreted that C: 14:00-16:00 is the parallel third session of the course from August 4th to August 14th.
If the inquiry is made without a screenshot, it can be answered normally.
Regarding this issue, we will review the solution with the product and R&D colleagues this afternoon.
Issue 2:
Course schedule for the Municipal Federation of Trade Unions campus.
We tested in the sandbox environment, and currently, we can see the Friday course schedule. We checked the history, and we are not sure if this change is due to you having optimized the knowledge base content.
Figure 1: Conversation effect in the sandbox environment:
Knowledge base search results:
When you tested, the results retrieved from the knowledge base search did not show the Friday time slot course schedule description in the same format, nor the description in the remarks. This added a certain degree of ambiguity to the model’s understanding, leading the model to refrain from outputting in order to ensure stability.
Would it be better to remove the note?
When using the “Look up knowledge base” action in a guide or skill, can you specify which document to query?
I have already optimized the knowledge base by adding the summer class discount information to the course descriptions, but issues still arise. For example, a student from No. 7 Middle School enrolling in the 2 PM class did not receive the 100-unit discount.
For the Xiwai Campus, no discount policy is set, yet the system responded to the customer indicating a bundled registration discount.
Hello, we have reviewed the configuration in your system space. Regarding today’s issues, here is a consolidated response:
-
“Would it be better to delete the remark explanation?” The core issue here is that the knowledge base retrieval during the original conversation returned a snippet regarding “course arrangements for the City Federation of Trade Unions campus” for the fall semester, which ended at the red box, as shown below. It was missing the Friday explanation in the same format. I see that your new knowledge base has been optimized, and sandbox testing shows no issues.
-
“When using the ‘Consult Knowledge Base’ action in guidance or skills, can I specify which document to query?” The current system does not yet support this feature. We have submitted this requirement to the product development team and are awaiting feedback on the scheduling.
-
“I have already optimized the knowledge base by adding discount information for summer classes to the course descriptions, but issues still occur. For example, students registering for the 2 p.m. class at No. 7 Middle School do not receive the 100 discount,” and “For the Xiwan campus, no discount policy is set, yet the response tells the customer there is a bundled registration discount.” From the backend, it appears this may be caused by a combination of two issues:
a) The “Quotation and Fee Calculation” skill is configured as shown below, leading to the understanding that the discount applies universally.
b) The second issue is that the current skill configuration requires more detailed description for logic involving the stacking of multiple discounts.
We will continue to address these points tomorrow. Once resolved, we will conduct further testing and provide you with an update. -
“In the sandbox test, the City Federation of Trade Unions course for the fall semester includes Friday, but the AI response does not.” As mentioned in our reply this morning, this issue stems from the customer’s image input recognition in the preceding context. We tried several solutions this afternoon but have not yet resolved it. We will continue to push forward tomorrow.
Hello! To address this issue, you can make the following adjustments.
You can modify the configuration from two perspectives:
First, add a “Source Information Priority” rule in the role settings. The core logic is: the knowledge base is the only trusted data source. Content from user-uploaded images/OCR can only serve as clues for user intent (i.e., “what the user wants to ask”), but cannot be used as factual evidence. When OCR results conflict with the knowledge base, prioritize the knowledge base.
Second, add a “Off-Process Task — Handling Image-Based Inquiries” rule under conditional compliance. When a user sends an image and inquires about courses, prices, or time slots, you must first call the knowledge base to search using the user’s text keywords. Confirm based on the results returned by the knowledge base, while using OCR only as auxiliary reference.
Reason for the previous issue: When customers send table-type images such as course schedules, the OCR may misinterpret the row and column relationships. For example, in this summer class schedule, the left side lists “Sessions 1–4,” and the right side lists “Time Slots A/B/C/D.” The actual relationship is that each session has four selectable time slots (a cross-combination relationship). However, the OCR might interpret this as a one-to-one binding: “Session 1 = Slot A,” “Session 2 = Slot B,” “Session 3 = Slot C.” Therefore, constraining the knowledge base as the only trusted data source can resolve the OCR recognition errors.
Hello! To address the issue of inaccurate price calculations, we recommend modifying the skill using the following approach.
For the skill: “Quotation and Fee Calculation,” the overall direction for modification is to shift from a logic that “directly answers the price” to one that follows a “standard calculation process + pre-output review.” This will make quotations more stable when dealing with combinations of multiple shifts, sessions, and participants.
The core idea is to first break down the class combination the customer wants to quote, then query the price and duration for each class separately, and finally calculate the discounts and total price uniformly. This prevents the model from jumping directly to the total price based on the user’s description, thereby reducing omissions and calculation errors. Additionally, add a pre-quotation review mechanism. Before providing the final response, have the model verify that class matching, price queries, discount checks, and duration summaries have all been completed. Only output the result to the parent after confirming everything is correct.
You can refer to the image below for guidance on making these modifications:
I went through the entire guide and the skills, and it seems there’s no place where this price calculation skill is referenced. If this skill isn’t referenced or called, will the AI still trigger it automatically, correct?
