llm_markdown_parser 0.1.0
llm_markdown_parser: ^0.1.0 copied to clipboard
Pure Dart Markdown for streaming LLM replies: incremental parsing, healing of half-typed syntax, plain text, think tags and server-sent events.
llm_markdown_parser #
Markdown for text that arrives in pieces, such as a streamed LLM reply. Pure Dart: it runs on a server, in a CLI, in an isolate or in Flutter.
This is the text half of llm_markdown, which adds the Flutter widget.
- Incremental parsing. Finished blocks are parsed once and kept. Only the last, open block is parsed again when text arrives.
- Healing.
**bold, an open code fence or a half-typed[link](https://...is closed before parsing, so it never shows as raw symbols. - Plain text of a reply, for copy, search, notifications or text-to-speech.
<think>splitting for models that put their reasoning first.- Server-sent events from a byte stream.
Usage #
import 'package:llm_markdown_parser/llm_markdown_parser.dart';
import 'package:markdown/markdown.dart' as md;
final parser = IncrementalMarkdownParser();
var text = '';
await for (final chunk in chunks) {
text += chunk;
parser.update(text, complete: false);
// parser.frozenBlocks never change again; parser.openNodes is the
// block still being written
}
parser.update(text, complete: true);
final html = md.renderToHtml([
for (final block in parser.frozenBlocks) ...block.nodes,
]);
Smaller pieces work on their own:
healIncompleteMarkdown('Some **bo'); // 'Some **bo**'
markdownToPlainText('# Hi\n\n**Bold** text'); // 'Hi\n\nBold text'
splitThinking('<think>plan</think>Answer', complete: true);
await for (final event in serverSentEvents(response.stream)) { ... }
The nodes are markdown nodes:
GitHub-flavored Markdown plus math ($...$, \[...\]), alerts,
<details>, citations ([1]), ==highlight==, ||spoiler|| and emoji.
MarkdownSyntax turns each of them on or off and takes your own syntaxes.
How it stays correct #
The text is cut only where nothing that comes later can change the block before it. The tests stream sample replies, a few thousand generated documents and every example of the CommonMark and GitHub specs, and check that the finished blocks equal a normal parse of the full text.
License #
MIT