Brand Knowledge Base (Product Evidence)
Upload manuals, price sheets, or pages — articles cite these facts; a wrong price gets blocked by the QC gate.
The brand knowledge base is WalkGEO's content evidence hub: turn product manuals, price sheets, and product pages into structured, checkable facts for AI. When you write articles, generated content cites only these facts — and any number that contradicts the knowledge base (a wrong price, for example) is blocked by the quality gate and never published.
Where to find it: Workspace → Brand Hub → the Knowledge base tab on the left.
Step 1: Feed knowledge (three ways)
| Method | How | Best for |
|---|---|---|
| Upload a file | Click "Upload file" and choose PDF / DOCX / MD / TXT | Manuals, price sheets, spec sheets (≤20MB, 20 documents per brand) |
| Import from URL | Click "Import from URL" and paste a product / FAQ / help-center page | Product info already on your website (the system fetches the article body; SSRF protection built in; only URLs you explicitly add are fetched) |
| Add text | Click "Add text" and paste markdown or a table directly | Day-to-day fact updates, e.g. "price changed to ¥299" |
Cold-start tip: start with a "minimal knowledge card" — 3–5 product facts, 1 price, 3 FAQs. Five minutes and you're live. With an empty knowledge base, everything behaves exactly as before — nothing changes in existing flows.
After import the system automatically chunks your content (spec and price tables stay whole, never split mid-row), classifies entries (pricing / spec / FAQ / feature / policy), and records sources and timestamps.
Step 2: Review and maintain (imports are not a black box)
Every document in the list carries a status badge:
- Ready: imported and citable
- Queued / Parsing: being processed
- Failed: with a written reason. For example "PDF has no text layer" — scanned price sheets can't yield reliable numbers, so re-enter key prices via Add text (by design: no OCR guessing, so wrong prices never enter the knowledge base)
Open any document to inspect its evidence chunks (kind + version + content preview):
- Disable / enable a single chunk to temporarily retire a fact without deleting the document
- Re-parse after the source file changes
- Delete removes the document and all its chunks permanently
Conflict adjudication
When the same product/spec has contradictory entries (e.g. website price ¥2,999 vs an old price sheet ¥2,599), a "Knowledge conflicts" panel appears at the top. Adjudicate promptly: click "Keep this" on the correct entry — the rest are superseded automatically. Unresolved conflicts make retrieval and quality checks pick randomly.
Step 3: Just write articles (no extra steps)
Once knowledge is in, every writing path automatically becomes evidence-backed:
- Article creation / AI automation: before generation, the most relevant evidence chunks are injected into the writing instructions. Product specs, prices, and feature claims must come from the knowledge base, and the draft records which entries were cited (traceable).
- Wrong numbers get blocked: any price/spec in the draft that contradicts the knowledge base → the quality gate fails → forced to draft, never published to your site.
- Scores benefit: the share of traceable product assertions feeds the article score's citation and trust dimensions.
- AI assertion checks: beyond exact number matching, an LLM pass extracts product assertions from the draft and judges each against the knowledge base (supported / contradicted / unsupported); verdicts are stored with the draft.
Topics also get a "KB-backed" badge
On the Article Creation page, AI topic generation prefers angles your knowledge base can support — the selected topic shows a "KB-backed" badge, meaning the article stands on checkable facts rather than model memory.
When knowledge changes (price-change scenario)
After a price change, enter the new price into the knowledge base. The system knows which drafts cited the old value (every citation stores a version snapshot):
- Call
kb_stale_report(below) to list every article that cited the old version - A later release will route refresh tasks into Issues automatically (manual confirmation by default — published articles are never rewritten without your approval)
AI Agent / MCP users
With the WalkGEO MCP installed, your agent gains 10 knowledge tools. Recommended workflow:
kb_doc_import_url(brand_id, "https://your-site/product-page") ← import knowledge
kb_doc_add_text(brand_id, title, content) ← or plain text
kb_search(brand_id, "Pro X3 price") ← self-check evidence before writing
compose_article(...) ← citations automatic, wrong price auto-blocked
kb_doc_list / kb_doc_get / kb_doc_reparse / kb_doc_delete
kb_conflicts_list → kb_conflicts_resolve ← adjudicate conflicts
kb_stale_report(brand_id) ← articles affected by a price change
FAQ
- "PDF has no text layer"? The file is a scan. The system won't OCR-guess numbers (wrong prices must never enter the knowledge base) — re-enter key prices and specs via Add text.
- Will my documents be public? No. The knowledge base is used only for your own article generation and quality checks; it is never exposed externally.
- How granular should a document be? Don't worry — chunking is automatic and tables stay whole; use Disable to retire noisy chunks.
- Quota: 20 documents per brand, 20MB each. When full, delete outdated price sheets first.