商店页面
边缘世界 RimWorld
全部 讨论 截图 艺术作品 实况直播 视频 创意工坊 新闻 指南 评测

边缘世界 RimWorld

商店页面
RimWorld >  创意工坊 > SNAC 的创意工坊 > RiMind > 讨论
RiMind
描述 讨论10 留言236 改动说明
taotao 9 月 9 日 上午 8:10
It seems the quota exhaustion issue I reported earlier is still present. This time, however, I have more definitive test results.
The Problem:
I am frequently encountering the [RiMind] All available API keys and models failed error message. When I check the logs, I see that every single API request is being rejected with a 429 Too Many Requests error code.
Crucially, this happens even under conditions of very low dialogue frequency, with only two colonists in my game. It is not an issue of exhausting my daily quota.
My Configuration:
I have configured 5 separate API keys from 5 different Google Cloud projects.
For each key, the mod is set to try 5 different models in sequence.
Detailed Observations & Analysis:
From the logs, I can see a very specific failure pattern:
Initial Failure: The very first API request for a dialogue often fails immediately with a 429 error.
Rapid-Fire Retries: Immediately after the first failure, within the same second, the mod attempts to call the next model in the list. This also gets a 429 error.
Cascade Failure: This process repeats at a very high speed (e.g., two requests per second). The mod cycles through all 5 models on the first API key, then immediately jumps to the second API key and repeats the rapid-fire process, and so on. This results in a cascade of 25 consecutive 429 errors in just a few seconds.
< >
正在显示第 1 - 4 条，共 4 条留言
taotao 9 月 9 日 上午 8:13 
To confirm the server-side behavior, I checked the Google Cloud Console for each of my 5 API keys. The API request logs show an identical pattern for every key: exactly 5 requests were made, and all 5 of them received a 429 error response from Google. This confirms the mod is successfully reaching the API but is being actively throttled.
#1
SNAC  [开发者] 9 月 9 日 上午 8:34 
Thank you for the report. We are actively testing various methods to track down the root cause of these persistent errors. As you've astutely noted, this doesn't appear to be a true 'quota' issue. It seems that the API returns this specific error code for some unknown reason when it's unable to process a request.

To analyze this, I've conducted a test where I received a 429 error, immediately changed the prompt, and resent the request. In some cases, I received a normal response. This suggests the cause might be a specific format or characteristic of the prompt being sent to the API server. However, since the same prompt can sometimes cause an error and other times work just fine, it's taking a considerable amount of time to investigate.

Therefore, as time allows, I am testing with various prompt configurations. If this issue cannot be resolved, we may have to make significant prompt adjustments, which could mean sacrificing some existing features.
#2
SNAC  [开发者] 9 月 9 日 上午 9:00 
引用自 taotao：
To confirm the server-side behavior, I checked the Google Cloud Console for each of my 5 API keys. The API request logs show an identical pattern for every key: exactly 5 requests were made, and all 5 of them received a 429 error response from Google. This confirms the mod is successfully reaching the API but is being actively throttled.
Thank you so much for doing the experiments yourself. Based on my own tests with four keys, if a single request from just one key gets a 429 error, all subsequent requests using the same prompt are rejected across every key. I'll use your findings to better analyze the exact cause. I'm sorry you're unable to use it right now.
#3
taotao 9 月 9 日 下午 9:23 
引用自 SNAC：
Thank you for the report. We are actively testing various methods to track down the root cause of these persistent errors. As you've astutely noted, this doesn't appear to be a true 'quota' issue. It seems that the API returns this specific error code for some unknown reason when it's unable to process a request.

To analyze this, I've conducted a test where I received a 429 error, immediately changed the prompt, and resent the request. In some cases, I received a normal response. This suggests the cause might be a specific format or characteristic of the prompt being sent to the API server. However, since the same prompt can sometimes cause an error and other times work just fine, it's taking a considerable amount of time to investigate.

Therefore, as time allows, I am testing with various prompt configurations. If this issue cannot be resolved, we may have to make significant prompt adjustments, which could mean sacrificing some existing features.

Thank you so much for your quick and detailed reply!
It sounds like a very tricky and complex issue to debug. Your theory about the prompt itself being a potential trigger is very insightful, and it certainly explains why the error can be so unpredictable.
I completely understand that this will take time to investigate. I'm more than happy to continue testing on my end and will let you know if I find anything new.
Thank you again for your hard work and transparent communication
#4
< >
正在显示第 1 - 4 条，共 4 条留言
每页显示数： 1530 50