Block 06 · Document AI Search · canonical demo

Find anything in 47 documents. In 4 seconds. With citations.

Upload contracts, bylaws, manuals, policies. The system reads them, embeds them, and answers natural-language questions in your employees' native language — every answer linked back to the source paragraph.

Example query

"what is RLHF and how does it differ from supervised fine-tuning?"

→ Returns a synthesized 2-paragraph answer with footnote citations to specific document pages. Click the footnote, the source PDF opens to that page with the relevant section highlighted.

Try the live demo Karpathy LLM Wiki — 47 docs indexed → https://pkm-wiki.evan-ratner.workers.dev/

What you upload

PDFs · DOCX · text · email exports. Most law firms / management companies have 200-2,000 docs. We process the lot in ~30 minutes.

What it does

Semantic search (not keyword) + Q&A synthesis + citation-back. Answers questions like "where is our force-majeure language?" in 4 seconds versus 20 minutes of paralegal time.

What it costs

$5,000 setup · $750/mo for a firm with ≤2,000 docs. Setup includes ingestion, taxonomy tuning, employee training session, and a custom voice profile if needed.

How to sell this block

  1. Get 5 public PDFs. Bylaws, FAQ, a sample engagement letter. Don't ask for confidential material on Day 1.
  2. Ingest. Build a hosted demo on THEIR docs. Password-gated subdomain. Send the URL during discovery.
  3. Type ONE question they can't easily answer. The synthesized answer + page-level citation lands in 4 seconds.
  4. Anchor on partner time. A $500/hr partner spends 30 min/wk answering "where is X?" That's $13k/yr in labor. The system pays for itself in 6 weeks.

Deploy doc search for your prospect

Sales runs /cafecito-blocks:docs <prospect-slug> with a folder of their public PDFs. The skill ingests them, builds a hosted demo at docs.cafecito-ai.com/<slug>/, drafts the cold-pitch.

Read the playbook GitHub mirror Live PKM Wiki →