Blog
Latest articles and updates
We swapped Gemini embeddings for a free 300M model on a laptop CPU. It found the same answers
Our search runs on Google's Gemini embeddings. We tried replacing them with small open models running on our own hardware, tested on 269 real searches our AI assistants ran, graded blind. Stock models fell short. One cheap trick closed the gap on finding answers.
We loaded 1,000 government PDFs into Google Drive and Warmhand. We lost on speed, fixed it, and ran it again
984 federal PDFs in 40 folders, the same 40 questions through the same Claude, graded blind. Warmhand was more accurate but twice as slow. We found why, made search about 10x faster, and ran the whole thing again.
We put our handbooks in Google Drive and ended up with the same answers at six times the reading
Three team handbooks in Google Drive and in Warmhand packs, 58 questions through the same Claude. Both got every real answer right. What each handed the model, where our ranking lost, and how each handled questions the handbook does not cover.
We put our sales handbook up against a Claude Project and ended up rebuilding search for long documents
The same 23 handbook files in a Claude Project and in a Warmhand pack, asked 43 blind questions through the same Claude. Both answered all 43. What it took to get there, and how the two differ.
How teams share what they know with AI today: an honest comparison
Google Drive, Notion, Obsidian, a shared text file, Claude and ChatGPT Projects, Guru and more: what each does well, where it stops, what it costs, and the problem Warmhand is trying to solve.
We benchmarked our packs search against Context7 and ended up rebuilding it
We benchmarked Warmhand docs packs against Context7 on six libraries with blind, held-out questions: how a tuned set fooled us, what we changed, and where each side now leads.