Build Knowledge Base From Docs

A business's real answers live in documents (policies, product sheets, FAQs, process guides). This skill makes them answerable everywhere TrustPager AI runs — the in-app assistant and voice agents — by loading them into AI Knowledge.

Step 1 — Convert each document to Markdown

For every file the operator points at:

python ~/.claude/bos-run.py tool markitdown_convert "<path>"

Handles PDF, Word, PowerPoint, Excel, HTML. If MarkItDown isn't installed, relay its install hint and stop. If a file converts empty (e.g. an image-only scan with no OCR text), flag it rather than loading a blank entry.

Step 2 — Split into clean knowledge entries + show the plan

AI Knowledge works best as focused entries, not one giant dump. From the Markdown, propose a split: one entry per topic / policy / FAQ, each with a clear title and the relevant content. Show the operator the proposed entries before writing:

From "Lending-policy-2026.pdf" I'd create 6 knowledge entries:
  • "Minimum deposit requirements"
  • "Acceptable income types"
  • "Self-employed applicants" ...
From "Product-sheet.docx": 3 entries. Total: 9 entries.
OK to load these into AI Knowledge?

Keep titles in the customer's own words. Don't invent content that isn't in the documents; don't merge unrelated topics into one entry.

Step 3 — Load into AI Knowledge (after confirmation)

Create the entries in TrustPager AI Knowledge (create_knowledge / ingest_knowledge_base_text, per the platform's knowledge tools). Batch them; stream progress. Respect the approval queue (202 = queued, surface it — knowledge/safeguards.md). Tag the source document on each entry so it's traceable and easy to refresh later.

Step 4 — Confirm + point at where it's used

✓ Loaded 9 entries into AI Knowledge from 2 documents.
  Your in-app assistant and voice agents will now answer from these.
  Re-run this when a policy changes to refresh the entries.

Hard rules

• ❌ Don't load empty or image-only conversions — flag them instead.
• ❌ Don't invent or merge content — entries reflect what the documents say.
• ❌ Don't write to AI Knowledge without showing the proposed entries first.
• ✅ Always convert through tools/markitdown_convert.py — no bespoke parsing.
• ✅ One focused entry per topic; titles in the customer's words; source tagged.

Output shape

The proposed entry list per document for confirmation, then a one-line load confirmation (how many entries, from how many docs) and where they're now used.