Knowledge bank ingest.

Every tenant gets two Qdrant collections, hard-isolated:

• company_kb — your own docs, FAQs, policies, scripts. Long-form. Voyage-3-large.
• customer_data — per-customer state, orders, tickets, plan. Shorter. Voyage-3.

Both are queried on every turn. The rule book can scope which one is consulted.

Dashboard → Knowledge bank → New artifact. Supported formats:

• Markdown (.md, .markdown)
• PDF (text-extracted; OCR pass-through for scanned PDFs)
• DOCX
• HTML / web pages (paste URL, we crawl one page; bulk URL list also works)
• Plain text (.txt)

Each artifact gets semantically chunked (500–1000 tokens per chunk, sentence boundaries respected), embedded with Voyage-3-large, stored with rich metadata (title, section, source URL, last updated).

If your docs live in a Git repo or an S3 bucket, point us at it. We poll on a cadence (default daily) and re-index on changes. Authentication via deploy key or S3 IAM role.

Shopify — OAuth install via the Shopify app. We sync customers, orders, returns, subscriptions on webhook events plus a nightly full pass.

Generic REST — Paste an endpoint + auth header. We poll on your schedule. Returns rows in your shape; we map fields to our Contact / Order / Ticket models.

SQL (JDBC / SSH) — Provide a read-only DB user + connection string. SSH tunnel supported. Run a discovery query, map columns, schedule incremental pulls.

CSV upload — Drag-and-drop. Schema-discovery wizard infers types and runs through a merge preview before commit. Best for one-off campaign uploads.

One customer, three data sources, three different identifiers — the orchestrator merges them into one Contact:

• Email + phone as primary keys (fuzzy-normalised: lowercase, E.164)
• External IDs preserved per source (so we can write-back to Shopify by their customer_id while writing back to your CRM by yours)
• Conflicts surfaced as merge candidates for you to confirm; we don't merge silently across high-confidence-but-not-certain matches

Before every LLM call, customer-data fields run through the redactor:

• Phone (E.164 + Indian regional + US 10-digit) → {{phone:N}}
• Email → {{email:N}}
• PAN, Aadhaar (India) → {{pan:N}}, {{aadhaar:N}}
• SSN (US) → {{ssn:N}}
• IFSC, bank account → {{ifsc:N}}, {{bank:N}}
• Money amount → {{money:N}}
• Address (multi-line postal) → {{addr:N}}

The model reasons over the tokens; we reinject originals before the customer hears or reads the reply. Logs keep redacted form by default; full-fidelity audit trail available to DPOs.

Default: top-5 chunks from company_kb + top-3 chunks from customer_data scoped to the current contact. Tuneable per rule book.

Re-rank pipeline runs over the initial recall — Voyage rerank-3 — to surface the paragraphs that actually answer the query, not just the closest neighbours.

Re-uploads replace the previous version. Old chunks are deleted; new chunks indexed. A version history is retained (per artifact) for rollback.

If you want to A/B test KB versions on different campaigns, scope the artifact to a workflow rather than a tenant.