Knowledge infra for your coding agents.
Traditional RAG is doomed. Period.
Featured
Featured Articles
-
Knowhere Can Now Plug Into Your Agent
Knowhere Notebook and MCP let agents in Cursor, Claude, and Codex search a shared cloud document Brain.
-
Does 90% of a RAG Project Have Nothing to Do With the Model?
Production RAG quality is mostly evaluation and data/state work—not model swaps.
-
Why RAG needs a World Model too
Document agents need persistent world state, not just retrieved observations.
Browse
All Articles
-
1,500+ Stars in a Month: What Knowhere Adds Beyond MinerU
Knowhere is post-parse agent memory—tree hierarchy, multimodal assets, graph, and agentic retrieval—not a MinerU replacement.
-
The Four Eras of OCR: From Manual Features to Multimodal Large Models, A History of Text…
A history of OCR from mechanical templates through deep learning to multimodal large models.
-
Is SkillOpt learning a skill or adjusting a prompt?
Reflecting on Microsoft SkillOpt: is editing skill.md real skill learning or prompt optimization?
-
SkillOpt: Building Agents Takes More Than Just Prompts and Fine-Tuning
An explainer of Microsoft Research’s SkillOpt: treat skill.md as trainable text parameters.
-
How to Build RAG in Harness Engineering
Applying Agent = Model + Harness to RAG failures rooted in ingestion and structure loss.
-
AI Hallucinations Aren’t Always the Model’s Fault
Flat fixed-window chunking causes many so-called hallucinations; structured document memory helps.
-
After months of work, our document parsing engine is now open source — built to kill AI…
Knowhere open-source launch: tree-like parsing, table fidelity, OpenClaw plugin, SaaS and self-host.
-
Why Raw Text Is the Wrong Data Layer for RAG (and What to Do Instead)
Knowhere API as structured ingestion: parse, semantic chunk, and ship RAG-ready manifests.
-
I Finally Cracked the Code on Why LLMs Fail at Real Work
Visa-consultancy narrative: SnapFill + Knowhere raises form-fill success from ~20% to ~90%.
Blog 