Knowledge base & connectors
What to upload, how it's chunked, how to keep it fresh.
The knowledge base is the corpus every suggestion is grounded in. If it isn't in here, RepCue won't say it: a card that cannot cite a retrieved passage is replaced by a battlecard fallback or nothing at all.
What to load
In rough order of payoff:
- Pricing: list prices, discount floors, packaging rules. Tables are kept whole; a pricing row is never split across chunks.
- Competitive battlecards: one per competitor, with trigger phrases.
- Objection handling / playbook: the words that work, in first person.
- Case studies: who, what, the number.
- Security & compliance FAQ: the questions that stall procurement.
- Implementation guide: timelines and what the customer has to do.
Supported sources
| Source | How | Notes |
|---|---|---|
| Upload | Drag-and-drop or picker in Knowledge | .pdf, .docx, .md, .txt, up to 15 MB |
| Paste | Add document → Paste text | Good for a single battlecard or FAQ |
| URL | Add document → From URL | Fetches the page, strips navigation, indexes the text |
| Notion | Connectors → Notion (OAuth) | Imports pages you select; re-sync pulls edits |
| Google Drive | Connectors → Google Drive (OAuth) | Docs, Slides and PDFs from chosen folders |
| Confluence | Connectors → Confluence (OAuth) | Spaces or individual pages |
| CLI / watch folder | python -m sotto_ingest ./kb --watch | For pipelines and bulk loads |
How documents become suggestions
- Parse into blocks: headings, paragraphs, tables.
- Chunk by heading, 400–800 tokens, 15% overlap within a section. A chunk never crosses a heading boundary and carries its heading path ("Pricing FY26 › Enterprise › Overage"); that path is what the card's why this footer shows.
- Embed with a production embedding model and store the vector next to a full-text index.
- Retrieve at call time with a hybrid of dense + keyword search fused by rank, under an 80 ms budget.
Document types
Tag each document as Pricing, Battlecard, Playbook, Case study, FAQ or Document. Types drive retrieval weighting (a pricing question prefers pricing chunks) and the HUD card kind.
Freshness
Each document shows fresh or stale based on age and type (pricing goes stale faster than case studies). Stale documents appear in the Knowledge header; Update re-uploads in place and keeps the citation history. Connector-backed documents re-sync on a schedule or on demand.
What reps see
Every card footer reads why this ▸ Document · Section · updated N days ago. Clicking it opens the passage. Reps can mark a card unhelpful → wrong fact; that signal shows on the document's performance panel so admins know which docs need attention.
Removing content
Deleting a document removes its chunks immediately; suggestions already shown on past calls keep their citation text for the audit trail but the link shows source removed.