Most AI summarizers suffer from a fundamental flaw: they replace a long page with a shorter, equally dense wall of text. Developer tamnt1707 has released Siftlume, a free, MIT-licensed Chrome extension that solves this by forcing the model to answer with UI components instead of paragraphs. The extension allows the LLM to select from a fixed library of blocks like charts, timelines, and verdict cards, ensuring outputs are scannable and structured rather than prose-heavy.
The Architecture of Constrained Generation
Siftlume operates on a five-step pipeline that prioritizes validation over freeform generation. First, it extracts content using specialized parsers like Readability for HTML, pdf.js for documents, and schema.org JSON-LD for structured data. The critical innovation lies in the prompt engineering: the system prompt is generated from a fixed library of OpenUI components. The model does not write HTML; it selects from a closed set of blocks including charts, timelines, SWOT analyses, and verdict cards. This ensures that every output has a known shape, preventing the layout clashes and styling nightmares common in LLM-generated HTML.
Context-Aware Extraction and Repair Loops
The extensionβs intelligence is not just in the LLM but in the extraction layer. It treats different content types with distinct strategies: YouTube videos are processed via captions with timestamped quotes, while GitHub Pull Requests are analyzed through the API to group changes by area and review state. A standout feature is the validation and repair loop. If the model emits invalid OpenUI Lang, the parser errors are fed back to the model for a single retry. This closed-loop system significantly increases the reliability of structured output compared to one-shot generation attempts.
Client-Side Privacy and BYOK Model Support
Unlike server-side summarization tools, Siftlume requires no backend infrastructure. The extension runs entirely in the browser, supporting a 'Bring Your Own Key' (BYOK) model. Users can configure API keys for OpenAI, Anthropic, Gemini, Groq, DeepSeek, or any OpenAI-compatible endpoint. For those prioritizing privacy or offline capability, the extension supports running small models locally via Ollama or LM Studio, leveraging the user's graphics chip. This architecture ensures that page content is only transmitted to the provider when explicitly requested, keeping API keys and browsing data strictly on the client side.
Key Takeaways
- Siftlume replaces LLM text generation with a fixed library of UI components to ensure scannable, structured outputs.
- The extension is fully client-side, supporting BYOK for major providers like OpenAI, Anthropic, and local models via Ollama.
- Content extraction is context-aware, using specialized parsers for PDFs, YouTube captions, and GitHub PRs.
- A validation and repair loop feeds parser errors back to the model to correct invalid component structures.
- The project is open-source under the MIT license, with code available on GitHub.
The Bottom Line
Siftlume proves that constraining an LLM's output format is often more valuable than increasing its generative freedom. By treating summarization as a component composition problem rather than a text generation problem, it delivers a genuinely useful tool for developers and researchers who need to scan information quickly.