Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
feat: Framework overhaul - object returns, JSX objects, streaming pat…
…tern, and context API (#36) Major overhaul of the framework based on my experience building out the first robust example (`HackerNewsAnalyzer`). This PR includes the following: - Support for object returns and nested JSX elements in objects - Adds a natural pattern for delineating between streaming and sync LLM output (`Component` vs `StreamComponent`) - A context API that can accumulate and track state across subgraphs (used to make streaming work without global state) - Add a robust `HackerNewsAnalyzer` example that summarizes posts from the front page (n=500) and runs sentiment analysis over all of the comments to product a report and a tweet in the style of Paul Graham. ## Object returns & nested JSX Add two features to streamline workflow authorship: 1. Support for returning objects instead of just Element Element[] and string 2. Add the ability to embed Elements inside of objects ### Direct object return - Components can now return plain objects directly instead of only JSX elements - Eliminates the need for wrapper components when combining multiple outputs - Example: `return { summary, commentAnalysis }` instead of needing to define another component that manipulates the output shape from tuple -> object ### Nested JSX Resolution in Objects - The execute function now recursively resolves JSX elements within returned objects - Enables natural composition where object properties can be JSX elements Example: ```tsx return { analyses: stories.map(story => ({ summary: <PostSummarizer story={story} />, commentAnalysis: <CommentsAnalyzer comments={story.comments} /> })) } ``` Together both of these changes eliminate the need for wrapper components, boilerplate, and reduces the awkwardness of having to deal with weakly typed tuples in fanout and parallel operations. ## Streaming Patterns LLM components need to support both (1) synchronous output -- the 99% use case -- and (2) streaming output -- the 1% output that is typically only used during the last step when output is being rendered to users via a UI. We want to optimize for sync output since this is the common use case. To support this we've added: 1. `gensx.StreamComponent`: a special kind of component that returns a standard formal with a plain text value and a stream to get the output 2. A standard `stream` prop available on all streaming components. This defaults to false, but can be set to true to trigger streaming for that component only. Further nested components revert back to sync invocation unless they specify `stream: true` as well. The final output can be streaming or not, depending on the types of the components that get returned and whether or not `stream` is set on those elements allowing you to mix/match if needed. This enables writing any component to either return a text value or a stream. Authors don't have to worry about supporting both formats. Consumers get the prompt value in the 99% use case or can set `stream: true` and get the expected behavior for the 1% use case. ## Context API A context API that can be used to accumulate and track state across different parts of the subgraph. Streaming is intrinsic to the framework, and the framework is very aware of the streaming context values. However, the general API should support other use cases in the future such as: - tracing - loggers - error boundaries - caching - configuration inheritence ## HN Example Adds a more robust real world example that analyzes the front page of HN: 1. pulls the top 500 pages, and filters down to text posts 2. summarizes the posts, analyzes comments for sentiment 4. feeds all of this into a report writer 5. rewrite the report in the voice of Paul Graham 6. writes a tweet about the post in the style of PG Powerful example in just 300 LOC! Co-authored-by: Jeremy Moseley <[email protected]>
- Loading branch information
