AI Documentation

SwiftChat runs on a provider-agnostic AI layer. Bring your own key or use the built-in Lovable AI Gateway.

Supported providers

| Provider | Models | Notes | |----------|--------|-------| | Lovable AI Gateway | GPT-4o, Gemini 2.5, Claude Sonnet | Zero setup, unified billing | | OpenAI | Any chat + embedding model | BYO key | | Anthropic | Claude 3.5+ | BYO key | | Google | Gemini | BYO key | | Local (Ollama) | Any | Requires reachable inference server |

Configure under Settings → AI Providers. Each tenant selects a default + fallbacks.

Features

• Reply Assistant — draft, translate, tone-shift, summarize
• Conversation Intelligence — sentiment, intent, next-best-action
• Lead Qualification — 0–100 score, refreshed on every inbound
• Sales Coaching — post-call recommendations, deal-risk flags
• Knowledge Base (RAG) — pgvector-backed, cited sources
• Global AI Search — natural-language search across the whole tenant

Knowledge Base

• Upload PDFs, docs, or connect URLs — chunked to 512 tokens, embedded, stored in kb_chunks
• Retrieval uses cosine similarity with a re-ranker
• Every AI answer includes source citations linking back to the chunk

Usage & cost controls

• Per-tenant token quotas and hard caps
• Detailed usage log in BI → AI Analytics
• Fallback provider kicks in on 5xx or rate-limit responses