Local Llama Workflows Hit Web Wall

This is mostly a retrieval and content-shaping problem, not a model problem. The winning setup is likely a hybrid pipeline that strips boilerplate first, then falls back to structure-aware extraction and selective summarization only when needed.

• –Reader APIs like Jina Reader and Firecrawl are good at article pages, but they break down on app-like docs and interactive product sites.
• –Docling-style conversion and similar parsers help reduce token bloat, but they still need page-type-specific fallbacks for tables, nav-heavy layouts, and embedded widgets.
• –Screenshot-to-VLM is a viable escape hatch when text extraction fails, but it is expensive in tokens and works best as a last resort.
• –A small local extraction model can compress pages before the main LLM sees them, but that adds orchestration overhead and another failure mode.
• –For local models with tighter context windows, section ranking and query-focused retrieval are usually more scalable than feeding whole pages end to end.