{"id":3385,"date":"2023-08-17T11:00:01","date_gmt":"2023-08-17T18:00:01","guid":{"rendered":"https:\/\/devblogs.microsoft.com\/surface-duo\/?p=3385"},"modified":"2024-01-03T16:22:11","modified_gmt":"2024-01-04T00:22:11","slug":"android-openai-chatgpt-14","status":"publish","type":"post","link":"https:\/\/devblogs.microsoft.com\/surface-duo\/android-openai-chatgpt-14\/","title":{"rendered":"Prompt engineering tips"},"content":{"rendered":"

\n Hello prompt engineers,\n<\/p>\n

\n We\u2019ve been sharing a lot of OpenAI content the last few months, and because each blog post typically focuses on a specific feature or API, there\u2019s often smaller learnings or discoveries that don\u2019t get mentioned or highlighted. \n<\/p>\n

\n In this blog we\u2019re sharing a few little tweaks that we discovered when creating LLM prompts for the samples we\u2019ve shared. \n<\/p>\n

\n Set the system prompt \n<\/h2>\n

\n The droidcon SF sessions demo<\/a> has a few different instructions in its system prompt, each for a specific purpose (explained below):\n<\/p>\n

\n Keep the chat focused\n<\/h3>\n

The first part of the system prompt was intended to prevent the chat from going off-topic:<\/p>\n

\"You will answer questions about the speakers and sessions at the droidcon SF conference.\"<\/pre>\n
\n  Unfortunately, in gpt-3.5<\/code> the system prompt isn\u2019t given much weight. With gpt-3.5<\/code> this instruction helps the model to answer with the name of the conference but doesn\u2019t really restrict the types of queries it will respond to. With gpt-4<\/code>, however, the model follows this instruction more closely, with responses like this:\n<\/p>\n
\n  \"Model\n<\/p>\n
\n  Provide grounding data\n<\/h3>\n
The next part of the system prompt helps the model to answer some basic questions:<\/p>\n
\"The conference is on June 8th and 9th, 2023 on the UCSF campus in Mission Bay. \r\nIt starts at 8am and finishes by 6pm.\"<\/pre>\n
\n  The conference date, times, and location were originally added to the system prompt to help it answer questions about those facts. However, once we added the sessionsByTime<\/code> function<\/a> call the model can also use this information to determine appropriate parameters for the function.\n<\/p>\n
\n  Instructions on how to respond\n<\/h3>\n
Another part of the system prompt helps the model to respond with useful information:<\/p>\n
\"When showing session information, always include the subject, speaker, location, and time. \r\nONLY show the description when responding about a single session.\"<\/pre>\n
\n  This addition to the system prompt makes the model\u2019s responses more useful. Without it, the model won\u2019t always return all the information about the session (in this case it\u2019s missing the times and locations):\n<\/p>\n
\n  \"Chat\n<\/p>\n
\n  But with<\/em> the instructions to always include the metadata, it\u2019s much more verbose:\n<\/p>\n
\n  \"Chat\n<\/p>\n
\n  Note that the completion request was always including all<\/em> the metadata; the change to the system prompt is the only difference between these two different outcomes.\n<\/p>\n
\n  Even error messages can be better engineered\n<\/h2>\n
\n  In the post about adding a function<\/a> to query droidcon sessions by time, an error message is returned when no sessions are found \u2013 a scenario that would typically happen before or after the event.\n<\/p>\n
\n  The first iteration of the “no sessions found\u201d error was this text:\n<\/p>\n
\"There are no sessions available\"<\/pre>\n
\n  The model interpreted this into responses like this:\n<\/p>\n
\n  \"Error\n<\/p>\n
\n  By changing the function\u2019s response to include a reference to the conference:\n<\/p>\n
\"There are no sessions available.\r\nRemind the user about the dates and times the conference is open.\"<\/pre>\n
\n  the model gives a much friendlier response:\n<\/p>\n
\n  \"Error\n<\/p>\n
\n  Note that the error message doesn\u2019t need to specifically reference the conference dates, the model will understand that from previous context (in this case the system prompt, as explained above). The model will also use the current date and time to craft the response and refer to whether the conference hasn\u2019t started yet or is already over.\n<\/p>\n
Read The Fancy Manual<\/h2>\n
\n  The first time we discussed OpenAI chat function calling \u2013 to support the current weather query<\/a> \u2013 we also updated the system prompt<\/a> to contain:\n<\/p>\n
\"Only use the functions you have been provided with.\"<\/pre>\n
\n  This was added at the suggestion of the OpenAI documentation for function calling<\/a>, which states:\n<\/p>\n
\n  \"Screenshot\n<\/p>\n
\n  Sometimes prompt tweaks don\u2019t have to come by trial-and-error, they can come straight from the documentation \ud83d\ude09 Of course the function calling API is fairly new, so the possibility of hallucinated function names might be fixed in a future version.\n<\/p>\n
Summarize but not what you think<\/h2>\n
\n  The first OpenAI-related post this year was about implementing a summarization feature in an HTML editor<\/a>. The prompt suggested was:\n<\/p>\n
\"Strip away extra words in the below html content and provide a clear message as html content\"<\/pre>\n
\n  which you can test out in ChatGPT<\/a> \u2013 it works great and returns shorter, summarized HTML as shown:\n<\/p>\n
\n  \"Screenshot\n<\/p>\n
\n  It wasn\u2019t explained in that post, but there is a reason why the prompt isn\u2019t just \"summarize the below HTML\"<\/code>, and that\u2019s because when the prompt is that simple, the model interprets it \u201cliterally\u201d, with the following result (you can try it yourself with this source<\/a>):\n<\/p>\n
\n  \"Screenshot\n<\/p>\n
\n  As you can see, the model summarizes the HTML document, literally explaining the structure<\/em> as well as the content. This was run using GPT-3.5, but GPT-4 returns a similar style of response. A simple \u201csummarize the below\u201d works fine on plaintext though, so the more literal interpretation only applies when there is more potential for different interpretations (i.e., when there is a format or structure present in addition to the content). \n<\/p>\n
Resources and feedback<\/h2>\n
\n  We\u2019d love your feedback on this post, including any tips or tricks you\u2019ve learning from playing around with ChatGPT prompts.\n<\/p>\n
\n  If you have any thoughts or questions, use the\u00a0feedback forum<\/a>\u00a0or message us on\u00a0Twitter @surfaceduodev<\/a>.\n<\/p>\n
\n  There will be no livestream this week, but you can check out the\u00a0archives on YouTube<\/a>.<\/p>\n","protected":false},"excerpt":{"rendered":"
Hello prompt engineers, We\u2019ve been sharing a lot of OpenAI content the last few months, and because each blog post typically focuses on a specific feature or API, there\u2019s often smaller learnings or discoveries that don\u2019t get mentioned or highlighted. In this blog we\u2019re sharing a few little tweaks that we discovered when creating LLM […]<\/p>\n","protected":false},"author":570,"featured_media":3407,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[741],"tags":[734,733],"class_list":["post-3385","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","tag-chatgpt","tag-openai"],"acf":[],"blog_post_summary":"
Hello prompt engineers, We\u2019ve been sharing a lot of OpenAI content the last few months, and because each blog post typically focuses on a specific feature or API, there\u2019s often smaller learnings or discoveries that don\u2019t get mentioned or highlighted. In this blog we\u2019re sharing a few little tweaks that we discovered when creating LLM […]<\/p>\n","_links":{"self":[{"href":"https:\/\/devblogs.microsoft.com\/surface-duo\/wp-json\/wp\/v2\/posts\/3385","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devblogs.microsoft.com\/surface-duo\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devblogs.microsoft.com\/surface-duo\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/surface-duo\/wp-json\/wp\/v2\/users\/570"}],"replies":[{"embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/surface-duo\/wp-json\/wp\/v2\/comments?post=3385"}],"version-history":[{"count":0,"href":"https:\/\/devblogs.microsoft.com\/surface-duo\/wp-json\/wp\/v2\/posts\/3385\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/surface-duo\/wp-json\/wp\/v2\/media\/3407"}],"wp:attachment":[{"href":"https:\/\/devblogs.microsoft.com\/surface-duo\/wp-json\/wp\/v2\/media?parent=3385"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/surface-duo\/wp-json\/wp\/v2\/categories?post=3385"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/surface-duo\/wp-json\/wp\/v2\/tags?post=3385"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}