Is there a general solution to this problem? I assume you can only start buffering tokens once you see a construct, for which there are continuations, that once completed, would lead to the previous text being rendered differently. Of course you don't want to keep buffering for too long, since this would defeat the purpose of streaming. And you never know if the potential construct will actually be generated. Also, the solution probably has to be more context sensitive. For example, within code blocks, you'll never want to render links for []() constructs.
EDIT: One library I found is https://github.com/thetarnav/streaming-markdown which seems to combine incremental parsing with optimistic rendering, which works good enough in practice, I guess.
There are a few things in our implementation that make a more general solution unnecessary. We only need the output to support a limited set of markdown which is typically text, bullet points, and links. So we don't need code blocks (yet).
However, the second thing (not mentioned in the post) is that we are not rendering the markdown to HTML on the server, so []() markdown is sent to the client as []() markdown, not converted into <a href=...>. So even if a []() type link exists in a code block, that text will still be sent to the client as []() text, only sent in a single chunk and perhaps with the link URL replaced. The client has its own library to render the markdown to HTML in React.
Also, the answers are typically short so even if OpenAI outputs some malformed markdown links, worst case is that we end up buffering more than we need to and the user experiences a pause after which the entire response is visible at once (the last step is to flush any buffered text to the client).
Yes. You can define a regex matching what you want, and every regex can be compiled into a state machine (https://en.wikipedia.org/wiki/Nondeterministic_finite_automa...). Then at each character you make a step in your state machine. You pause the output while the regex is not matching.
fun read, its weird interacting with chatgpt around markdown sometimes.
it formats its own stuff with markdown, so if i ask it for markdown but dont explicitly specify a downloadable file, it will produce valid markdown up to where it conflicts with its own markdown, and then it gets choppy and chunked.
its an issue of my prompting is what im sure some customer service rep would be told to tell me :p because theres money to be made in courses for prompting skills perhaps, idk. (cynical view).
sure is enjoyable to struggle together with the AI to format its responses correctly :'D
Could this result in edge cases with [ where due to some misformatting or intentional syntax that looks like the start of a markdown link, the entire response is hidden from the user?
(This comment when subjected to this processing could look like: "Could this result in edge cases with ")
If you buffer starting with the ( character, then you'd still send the [text] part of the link, and worst case is that with no matching ) character to close the link, you end up buffering the remainder of the response. Even still, the last step is "flush any buffered text to the client", so the remainder of the response will be transmitted eventually in a single chunk.
There are some easy wins that could improve this further: line endings within links are generally not valid markdown, so if the code ever sees \n then just flush buffered text to the client and reset the state to TEXT.
I do something like this too because links in emails are insanely long. It's worse in marketing emails. So I shorten the links to save on tokens and expand them again when I get the response back from the LLM.
Is there a general solution to this problem? I assume you can only start buffering tokens once you see a construct, for which there are continuations, that once completed, would lead to the previous text being rendered differently. Of course you don't want to keep buffering for too long, since this would defeat the purpose of streaming. And you never know if the potential construct will actually be generated. Also, the solution probably has to be more context sensitive. For example, within code blocks, you'll never want to render links for []() constructs.
EDIT: One library I found is https://github.com/thetarnav/streaming-markdown which seems to combine incremental parsing with optimistic rendering, which works good enough in practice, I guess.
There are a few things in our implementation that make a more general solution unnecessary. We only need the output to support a limited set of markdown which is typically text, bullet points, and links. So we don't need code blocks (yet).
However, the second thing (not mentioned in the post) is that we are not rendering the markdown to HTML on the server, so []() markdown is sent to the client as []() markdown, not converted into <a href=...>. So even if a []() type link exists in a code block, that text will still be sent to the client as []() text, only sent in a single chunk and perhaps with the link URL replaced. The client has its own library to render the markdown to HTML in React.
Also, the answers are typically short so even if OpenAI outputs some malformed markdown links, worst case is that we end up buffering more than we need to and the user experiences a pause after which the entire response is visible at once (the last step is to flush any buffered text to the client).
Yes. You can define a regex matching what you want, and every regex can be compiled into a state machine (https://en.wikipedia.org/wiki/Nondeterministic_finite_automa...). Then at each character you make a step in your state machine. You pause the output while the regex is not matching.
This exact problem is why I wrote Streamdown https://github.com/day50-dev/Streamdown
Almost every model has a slight but meaningfully different opinion on what markdown is and how creative they can be with it.
Doing it well is a non-trivial problem.
Generating simple HTML instead of markdown would have been a solution. But I guess that ship has sailed.
fun read, its weird interacting with chatgpt around markdown sometimes.
it formats its own stuff with markdown, so if i ask it for markdown but dont explicitly specify a downloadable file, it will produce valid markdown up to where it conflicts with its own markdown, and then it gets choppy and chunked.
its an issue of my prompting is what im sure some customer service rep would be told to tell me :p because theres money to be made in courses for prompting skills perhaps, idk. (cynical view).
sure is enjoyable to struggle together with the AI to format its responses correctly :'D
You can ask for it to put the markdown in a codeblock. It works well for me. It also works with latex.
Could this result in edge cases with [ where due to some misformatting or intentional syntax that looks like the start of a markdown link, the entire response is hidden from the user?
(This comment when subjected to this processing could look like: "Could this result in edge cases with ")
If you buffer starting with the ( character, then you'd still send the [text] part of the link, and worst case is that with no matching ) character to close the link, you end up buffering the remainder of the response. Even still, the last step is "flush any buffered text to the client", so the remainder of the response will be transmitted eventually in a single chunk.
There are some easy wins that could improve this further: line endings within links are generally not valid markdown, so if the code ever sees \n then just flush buffered text to the client and reset the state to TEXT.
I do something like this too because links in emails are insanely long. It's worse in marketing emails. So I shorten the links to save on tokens and expand them again when I get the response back from the LLM.