[Question/Suggestion] Regarding API Call Notebook Optimization and Other Issues

OMate
1

Since the beginning of April, I have been closely following every update and learning new features of Omate! However, after this update, I encountered an issue where the notebook cannot be used normally, and I would like to seek help from Esey. I have also accumulated some minor issues and suggestions:

Device Information: Vivo x100s pro Version: 1.9.1

I. Software Issue Consultation

  1. Notebook Invocation Failure

After the 1.9.1 update, the previously smooth intermediary API can no longer invoke the notebook, resulting in a 400 error, indicating a tool invocation issue. After disabling the notebook feature, the AI assistant provides an explanation of the error content based on the current prompt and API output as shown in Figure 2.

Since the API could always invoke the notebook tool normally before the update, and this update mentioned optimization modifications to the notebook tool, I wonder if there is a bug… Testing with the official key of deepseek still allows invocation, but I am not sure what happened with the intermediary API.

Another related suggestion is that most players still use the intermediary more… I have recommended the use of the notebook to many friends (I use it to update characters, environments, and hierarchical memory), but most people gave up due to API limitations. Among the dozen or so APIs I use, only 3 can invoke the notebook (only one can do so stably). I consulted the intermediary provider, who said they support the Agent feature, so it might be an optimization issue with the Omate software. I wonder if the compatibility of the notebook Agent can be further optimized? This would better promote the “notebook,” a unique and infinitely potential tool of Omate.

  1. 998045ab-dd9b-4882-8f5d-a0cb0eadb532

    e4cd11750150327682747816cbfd990d

  2. Calendar Schedule Creation Issue with Agent Tool

Although an optimization was done once to add a calendar ID tool, the calendar schedule creation feature still cannot be used on this phone. When initially invoked, Omate requests permission to invoke the calendar schedule, but even after authorization, it still shows that the calendar name cannot be found, or there is a garbled error. I hope it can be further optimized, as companion users really want a small AI assistant that can interact with their daily lives! QUQ

  1. Software Stability Issue:

When too many messages are sent in the same conversation, it causes a crash. Specifically, under the last version of the 1.8 series, if the number of chats in the same conversation exceeds 1900, sending a message often results in a 10-second lag, and touching the screen may cause a crash. I personally checked that there is sufficient remaining memory on the device (about 200G), and no other software experiences crashes. After starting a new topic, this issue is resolved, and no more crashes occur. Additionally, when the number of chats exceeds three or four hundred in the same conversation, the AI response repetition rate soars, although the context count remains stable at 50. It seems there is an inherent instability when there are many chats in a conversation, and I hope there is an opportunity for optimization.
4. After the 1.9.1 update, I found that permanent memory search and addition both failed. The previously established index was also cleared and disappeared.

image

  1. Abnormal Log Issue:

The inference log continuously shows errors related to a deleted knowledge base (failed to search knowledge base “xxx某已删除知识库.txtxxx”, error message: Exception: Failed to read vectorized content: Exception: Vectorized file does not exist), which does not directly affect core functions, but the long-term existence of such anomalies may hide performance risks. It is recommended to investigate and fix this.

II. Function Optimization Suggestions

  1. Enhance Real-Time Time Function

As a heavy user of daily companion functions, the accuracy of real-time time is crucial. However, even with multiple prompt words, the AI still cannot stably respond to real-time time demands. I wonder if it can be suggested to add a forced trigger mechanism to ensure the function is stable and effective.

  1. Optimize Voice Reading Filter Function:

Currently, voice reading seems to only filter content within parentheses, and it reads symbols like >, -, and regular expressions, affecting the user experience. I hope a custom filter setting can be added, allowing users to independently choose which symbols or text rules to block, or at least filter out the reading of symbols.

These are the issues encountered for now, hoping for guidance, and thank you for reading!

2

  • As for the time, you can inject it into the prompt: The current time is {{year}}{{month}}{{day}}{{hour}}{{minute}}{{second}}. Use similar prompts, set the depth to 0, and choose the user for the role field, which will be more effective.

  • I am MIUI, and the default calendar ID is null, so it’s not usable, and I’ve given up :joy:

  • Previously, reading the notepad would lag, and the larger the text, the slower it got, sometimes causing a loop freeze. The new version of the official deepseek has improved a lot, but it’s still not recommended to store large amounts of text in the notepad for data exchange.

3
  1. For questions about the transit API, please refer to: 【建议/求助】关于gcli2API配置填写问题求助，以及几点优化建议 - #3 by gbye134340_89335 2. From the error screenshot, the AI-generated fields are not fields of our built-in tools, which may be due to AI malfunction or some inferior reverse-engineered interfaces. The [Tool Test] page in the menu at the top right of the chat page can be used for convenient testing. 3. We currently use the Android standard interface to obtain the calendar ID and have tested it on our test machines. Some manufacturers may have different interfaces, but we do not have testing equipment. Since there are not many users, we do not plan to purchase equipment for optimization in the short term. 4. The memory you understand is storage, not the actual runtime memory used. One area that can be optimized is to disable permanent memory. Other areas cannot be optimized, which is the main reason we were reluctant to add regex and HTML rendering before. 5. b212 upgraded the underlying implementation of full-text search (fts5), and for upgraded users, errors may occur due to the presence of old format data tables. This issue will be fixed in b213 later. 6. The issue with the knowledge base logs has been fixed, but errors will still occur for old role cards. 7. Voice reading does not use a filtering mechanism but instead has special marker symbols set. Set it in the role - [Read Text Content] section.
4

Thank you for the reply!

- There is a time prompt for the prompt words, but I chose the system. Indeed, using user might be more effective…

- I don’t know when the calendar will be usable… It’s been several iterations already (frustrated).

- The free tier official key for Gemini used to call notebooks very stably, but I don’t know why it can’t read long-term memory and knowledge base, so I had to give up… The APIs I used before didn’t work well. Recently, I switched to a new one and used it for almost a week. It’s very smooth and reads and writes very quickly. It also completed detailed diaries for 6-7 consecutive days (about 1000 words each day) very well, and it’s pay-as-you-go… I was very happy, but now with the update, I’m heartbroken. The only API left usable is the deepseek official key… It’s really frustrating.

5

Long-term memory and knowledge base are completely local and part of the context, not related to the agent function, meaning they are unrelated to the AI model. You should check if it’s limited by the maximum token limit.

The notepad tool requires the model’s toolcall, and there are quite significant differences between each model.

6

Thank you, Teacher Easy, for your reply! However, I feel very regretful and sad that I still cannot resolve the issue that appeared after this update.

- Regarding the API issue, I also read the post you recommended, but since several APIs I tested could be used continuously, stably, and smoothly before the update, they immediately reported a 400 error after the update. All other transit APIs currently tested are also unusable, only the official deepseek is available. I understand the moral significance of opposing transit, but the vast majority of users still use transit. I don’t ask for more compatibility, even if it is just consistent with the previous version? Such strict restrictions will greatly affect the promotion of users of functions like the notepad agent, wasting a potential feature, which is really distressing. I wonder if there is any plan, or if a related feature has already been released (but I didn’t notice), to configure the API separately for calling the notepad function?

- The calendar issue has been mentioned by many Android players as having problems with usability, and it is certainly not an isolated case. However, due to no one using it, there is no experience to correct it, so it can only be abandoned…

- As for memory, I also checked the running memory. The phone has 36G, and at that time, to rule out hardware issues, I also cleared other applications in the background, but there is still a problem of crashing when the current conversation chat rounds are too many. I only use one character card… I understand that it shouldn’t take up much memory? I don’t quite understand. After switching to a new topic, there is no crashing issue when the permanent memory still retains the original 1900 chat records.

- As for other things, I look forward to the updated version. I hope Omate gets better and better, and I also hope to promote it to more friends as a loyal player, so that everyone can apply the rich features.

7

Claude’s relay on my side is stable, but I can’t use the notebook… The official calls for Deep and Gemini’s notebook tools are very smooth. I really don’t understand why Gemini’s official key can’t use long-term memory and knowledge base, maybe it’s the token limit of the free tier (150,000 tokens). The context I provided is 600,000 tokens.

However, other relays for Gemini work very well (except for the difficulty or errors in calling the notebook). But now there is no free tier, the official key is dead… Relays can no longer call the notebook, some have issues with toolcall responses, and some directly report error 400. After the update at noon, I’ve been in a state of helplessness with various ineffective attempts to salvage the situation, feeling quite frustrated…

8

Yes, the part that exceeds the token limit is not visible to the AI. The AI can only respond based on this small visible portion.

What you mentioned sounds like the toolcall format of Omate changed after this update, but it shouldn’t have changed. Why don’t you try reverting to the old version?

9

It is very likely? After returning to AI assistant mode, it was reported that the excuse format was incorrect, and an attempt is being made to revert to the old version… I was originally excited to update and see how the memory-linked notebook works, but the notebook ended up being unusable orz. I just backed up, and I was a bit nervous about uninstalling and reinstalling for the first time… Thanks for the suggestion!

10

I would like to propose an optimization suggestion to Teacher Fangtang through this post, as I forgot to mention it in my own post before…

Regarding the red error box in the screenshot of the first point of the original post, could it be changed to disappear only after manual confirmation? Sometimes the screen is filled with text that disappears too quickly to read. It would be best if it supported text copying functionality, as copying it to ask AI can greatly improve troubleshooting efficiency.

11

Add the next version (b213)

12

I found the cause of the calendar issue. The plugin output was corrupted during the Android packaging, but the test environment was fine, no wonder I couldn’t detect it :joy:.

b215 will fix this issue. @YeZip @Dan_Lu

13

This topic was automatically closed 3 hours after the last reply. New replies are no longer allowed.

14

Related updates.