This browser is no longer supported.
Upgrade to Microsoft Edge to take advantage of the latest features, security updates, and technical support.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(179.20000000000002, 33%, 27%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EC%3C/text%3E%3C/svg%3E)
Hi all. We have just finished integrating Azure Open AI Service into our blog content management system. Given the following prompt: "I Want You To Act As A Very Proficient Tech Writer Who Writes Fluently In English. Write a 2000-word 100% Unique article in English that covers COMPANY NAME being accepted onto the Microsoft for Startup Founders Program. Write The article In Your Own Words. Use detailed paragraphs that engage the reader. Use An Professional & Authoriative Tone. End with a conclusion paragraph and 5 unique FAQs After The Conclusion. Context for the article includes: - COMPANY NAME joined in 2023 receiving a 25,000 USD sponsorship\r\n- The Microsoft Sponsorship has gone towards the development costs and resources for two of our AI products: PRODUCT 1, and PRODUCT 2" Via the API we received a response once which was garbage and seems to have completely ignored the settings provided to it:
".\r\nThese are the products that ultimately made COMPANY NAME stand out in Microsoft's eyes and we're excited to bring them to market soon. Finally yet importantly , google keyword “PRODUCT 1” for search engine optimised content.\n\nAugust 9th, 2021|Digital Marketing|\n\nPrevious\n\nInfographic - Digital Advertising Trends [2021]\n\nNext\n\COMPANY NAME 's Yeastar Cloud PBX Solutions brings tremendous savings for your company.\n\nRecent Posts\n\n- The Technical basis of modern-day SEO\n- Improving Local Rankings With GMB In Five Simple Steps\n- Top benefits small business owners will receive by leveraging Google My Business Listings.\n- Maximizing Your SEO Results Through Social Media Platforms\n- Power Up Your Online Presence: Secrets To Boosting Local Search Rankings For HVAC Companies & Contractors This Summer Season & Beyond!"
Calls to the API constantly fail with no error. We decided to try in studio and the response is even worse, and the studio keeps freezing and crashing. I tried to copy the output of Studio to show you what it is doing but it is completely frozen every time we try to generate. We have been using ChatGPT to write articles for months and have never seen anything as bad as this. Can anyone let us know why this is so bad? It is not what we were expecting at all. Model is gpt3-turbo because 4 is not available in west-europe (If you could advise when it will be that would be great) Max tokens set to 4000.
Forgot to mention using completion API.
This is an example of studio using chatbot as we are trying to find why it is so bad so trying different methods:
Prompt: I Want You To Act As A Very Proficient Tech Writer Who Writes Fluently English. Write a 2000-word 100% Unique article in English that covers An article about COMPANY NAME being accepted onto the Microsoft for Startup Founders Program. Write The article In Your Own Words Rather Than Copying From Other Sources. Use detailed paragraphs that engage the reader. Use An Professional & Authoriative Tone. End with a conclusion paragraph and 5 unique FAQs After The Conclusion. Context for the article includes: - COMPANY NAME joined in 2023 receiving a 25,000 USD sponsorship\r\n- The Microsoft Sponsorship has gone towards the development costs and resources for two of our AI products: PRODUCT 1, and PRODUCT 2.\r\n Resposne: PRODUCT 1 is a Machine Learning-Powered eCommerce Analytics Tool. \r\n - PRODUCT 2is a Can Automated Facebook Messenger Bot. \r\n - Innov8 Is Eager To Work With Microsoft To Develop More Machine Learning-Powered Products And Solutions And May Apply For The Next Cohort Which Will Begin Next Year. Article Length: 2,000 words Introduce Yourself First name Last name E-mail address Phone number Professional title Location Resume/CV Attach Resume/CV Here Upload requirements Portfolio To apply for this role, you will need to provide samples of your writing that demonstrate your skills in writing fluently in English on a professional level. You may provide links to your portfolio, and online profiles that demonstrate your work. Please provide us with any online profiles that you have that demonstrate your work. Please provide us with any links to your most recent work that demonstrates your writing ability. Additional Questions On a scale of 1 to 10, how proficient are you at writing in English? Do you have any experience working in corporate environments? If so, please describe. How do you manage a writing project or an article from start to finish? Please describe your process and how you stay organized and
Ah now our comments are dissapearing?
This is a response from Studio. We tried chatbot instead of completions to see if it was better but it is still outputting garbage. Same prompt as above.
Another example response from API:
\r\n- COMPANY NAME has integrated with Microsoft Teams and Schedule an interview feature on LinkedIn as part of their client services\r\n- COMPANY NAME uses SharePoint and Azure, Azure Active Directory to ensure optimal data safety for their clients. \r\n\r\nThis Article will be used for SEO purposes on our site so must follow or at least try to follow Google's guidelines in terms of writing style etc.\n It is completely making stuff up and not doing what requested. The same prompt in the chat app chat.openai.com works flawlessly.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(214.4, 74%, 30%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EAM%3C/text%3E%3C/svg%3E)
Innov8 Greetings!
I understand your frustration with the AI model generating irrelevant or inaccurate content. Unfortunately, AI models may sometimes generate irrelevant or inaccurate content.
However, there are some steps you can take to improve the quality of the generated content:
Also, check Learn how to work with the GPT-35-Turbo and GPT-4 models and Recommended settings and see if that helps.
Do let me know if you have any further queries.
Hi thanks for the reply. We resolved this by using chat, we are unable to use GPT-4 as it is not available in our region. We have a massive problem with reliability. We specifically instruct the model to return a response in valid json, we provide the exact format in both words and an example, but it constantly ignores this. We have provided the specific instructions as system prompts as well as user prompts. This is a real deal breaker, we have integrated it into so many of our systems now and they are constantly failing due to this. Any suggestions? To clarify, we are very very specific with our prompts and provide examples of valid JSON.
We have now switched to GPT4 and the responses are even worse, systems that were working a lot better just fail. This would be the end of the whole line of products we are building, is anyone able to provide feedback on this, we pass clear system and user instructions with the format of the expected json in words and in example. We have swapped to functions. It works for one call and the next call has returned null for over 2 hours, no response, no error, just null. We switched to GPT3.5 turbo 1106, GPT4 is actually terrible, never experienced as much pain as that :) Our soultion seems to be a lot better now with GPT3.5 and functions.
Ok well 6 days developing real-world systems with it has made us realise it is a completely unreliable and unusable platform if you expect to build systems around it. Systems will work one day then fail completely for hours. GPT4 is a lot worse than 3.5 turbo, 3.5 turbo will mostly listen to what it is told and using functions you can mostly get back what you need. One of our systems we have working well in 3.5 turbo, we switch the model to GPt4 and it instantly fails. All prompts are short, clear and concise providing the structure of the function in addition to JSON example of the expected response and description in text, it is not possible to create a stable system using this technology. Now we have functions randomly returning empty responses.
Innov8 This issue needs further investigation to find out the root cause. I would request you to file a support request @ https://aka.ms/azsupt for a deeper investigation and immediate assistance on this issue.
AshokPeddakotla-MSFT when I get some time I will do. To clarify a bit more. The article generation we mentioned earlier. We have now broken it down into sections, so the content for under each h2 is a single call, the content under each h3 is a single call etc. With GPT 35 turbo it works well but the model is not so good, we constantly have to modify it alot manually. When we switch the model to GPT 4, it just fails with no response. Usually on the first h3.
This is a shame because we have a system using it that literally runs our Facebook page, it is an FB app and it replies to every comment made on our page to increase engagement and promote conversations, and it also handles all messages coming to the page, it uses GPT4 and it is out of this world. (See this thread from tonight https://www.facebook.com/Innov8DigitalMediaTech/posts/pfbid0U7psaKH1rNESECbScDgfY2cN1EvhiUukzK4nuoL7hLYjk44XBsQSHLRST9HMEQr5l )
This is all part of one huge system that automates our content creation and marketing, but with GPT4 we are completely unable to use it for the content creation where GPT3 does it with no problems. When it fails it just returns NULL there is literally zero response from the api call, we log every response in a text file and when it fails in this case the file is empty.
I cannot create a request, there is no option to select Azure Open AI.
Ok despite documentation we found which stated otherwise. Completions in the way were trying at first does not seem to be the best way. Sending a request to /chat/completions with system and user messages seems to work well.
It is continuing to complete make up people and facts though, any way to stop this?