Knowledge bank

Drop in your docs. Every call learns at once.

Upload your policies, FAQs, scripts, PDFs. Connect Shopify, paste a REST endpoint, mount a SQL tunnel, drop a CSV. Voyage-3 embeds it, Qdrant stores it, every call queries it live. One source of truth, one tone, every campaign.

Company KB

Markdown, PDF, DOCX, web pages. Chunked semantically. Embedded with Voyage-3-large. Re-rank pipeline so the right paragraph wins, not just the closest neighbour.

Upload via dashboard or bulk-sync a Git repo / S3 bucket
Re-indexes on change; version-aware retrieval
Per-tenant Qdrant collection — your KB doesn't bleed into anyone else's RAG

Customer data

Orders, tickets, subscriptions, account state. Pulled live or synced on a schedule. Identity-resolved across sources so one customer = one Contact, not three.

Shopify OAuth · Generic REST · SQL (JDBC / SSH) · CSV upload
Fuzzy identity match with tenant-confirmed merges
Incremental sync via webhook or polling

PII-safe by default

Phone, email, PAN, Aadhaar, SSN, IFSC, bank account, money amount, address — redacted before the LLM ever sees them. Patterns ported from production code, not a workshop slide. Compliance is wiring.

"Customer phone +91-98765-43210, PAN ABCDE1234F, owes ₹4,500" → "Customer phone {{phone:1}}, PAN {{pan:1}}, owes {{money:1}}"

Original values are reinjected after the LLM responds. The model reasons over tokenised placeholders, your customer sees their real data. Logs and audit trails keep the redacted form by default.

Retrieval per turn

Every turn — voice or text — runs a retrieval. Top-k chunks from your company KB, plus the customer's own data (last order, open ticket, last call summary). The relevant context lands in the system prompt for that turn, not the whole document corpus.

You can scope retrieval per workflow. A renewal-nudge workflow might query "renewal policy" + "this customer's plan." A complaint-resolution workflow might query "refund policy" + "this customer's recent orders." The rule book decides.

Bring your own embeddings?

Not today. Voyage-3 and Voyage-3-large are the only embeddings we run. If you have a strong case for swapping in OpenAI, Cohere, or a self-hosted model, tell us at hello@callfunnel.ai. We're open to it; we just haven't hit a tenant who needed it yet.