Platform
AI Assistant
Talki's AI Assistant answers customer questions automatically using a knowledge base built from your website, documentation, and manual Q&A entries. Agents get one-click summaries of every conversation.
How it works
You build a knowledge base from your website, docs, PDFs, or manual Q&A entries
Customer sends a message — AI searches the knowledge base for a matching answer
In Suggest mode: AI drafts a reply for the agent to review
In AutoReply mode: AI sends high-confidence answers automatically
Low-confidence questions are flagged for human review
Building the knowledge base
The AI only answers questions it can find evidence for in your knowledge base. The more complete your knowledge base, the higher the accuracy. Three source types are supported — use any combination.
# Knowledge base sources (configured in Admin → AI → Knowledge base)
#
# Option A: Website crawler
# Enter your domain — Talki crawls public pages and indexes them.
# Pages crawled: /docs, /help, /faq, /pricing, /features, ...
# Re-crawl: manual or on schedule (daily/weekly)
#
# Option B: Manual entries
# Add Q&A pairs directly:
# Question: "How do I cancel my subscription?"
# Answer: "You can cancel anytime from Account → Billing → Cancel plan."
#
# Option C: File upload
# Upload PDF, DOCX, or TXT files (max 10 MB each).
# Good for: internal manuals, product specs, policy documents.Website crawler
The fastest way to get started. Talki crawls your public website and indexes all text content. Re-crawl manually or on a schedule to keep the knowledge base fresh after updates.
Manual Q&A
Add specific question-answer pairs for edge cases or internal knowledge that isn't on your public website. Useful for pricing exceptions, policy details, or team-specific procedures.
File upload
Upload PDFs, Word documents, or plain text files. Good for product manuals, spec sheets, or onboarding guides.
AutoPilot
AutoPilot controls whether AI replies are sent automatically or shown as drafts for agent review. Configure it in Admin → AI → AutoPilot.
# AutoPilot behavior (Admin → AI → AutoPilot)
#
# Mode: Suggest (default)
# AI drafts a reply for the agent to review before sending.
# Agent sees the draft pre-filled in the reply box — one click to send.
#
# Mode: AutoReply
# AI sends the reply automatically without agent review.
# A confidence threshold controls when AI replies vs. escalates:
# High confidence → reply sent automatically
# Low confidence → ticket flagged for human review
#
# Escalation keywords (bypass AutoPilot entirely):
# "refund", "legal", "urgent", "ASAP", "angry" — customizable list| Mode | Behavior | Best for |
|---|---|---|
| Off | AI features disabled. No drafts, no auto-replies. | Teams not ready for AI assistance |
| Suggest | AI drafts a reply — agent reviews and sends. | Most teams. Keeps humans in the loop. |
| AutoReply | AI sends high-confidence answers automatically. | High-volume, repeatable questions |
⚠AutoReply recommendation
Conversation summary
Every ticket has a Summarize button in the header. Click it to generate a concise summary of the conversation — useful for handing off tickets between agents, getting context before jumping into a long thread, or escalating to a senior rep.
# One-click conversation summary (available on every ticket)
#
# Summarizes:
# - What the customer issue is
# - Current status and what's been tried
# - Recommended next action for the agent
#
# Powered by Gemini. Summary is generated on demand — click the
# "Summarize" button in the ticket header.Confidence & escalation
Every AI answer has a confidence score (0–1). You can set the threshold in Admin → AI → Confidence threshold. Answers below the threshold are never sent automatically and are flagged for human review regardless of AutoPilot mode.
High confidence (above threshold) — AI answer is sent (AutoReply) or shown as a pre-filled draft (Suggest).
Low confidence (below threshold) — Ticket tagged needs-human and routed to the unassigned queue for agent review.
Escalation keywords — Configurable list of words that bypass AutoPilot entirely (e.g. "refund", "legal", "cancel").
AI answer via REST API
You can query the AI knowledge base directly via the REST API — useful for building custom chat interfaces, in-app search, or enriching your own support tools.
# AI endpoints (org-scoped, requires auth)
# Generate a reply suggestion for a ticket
POST /v1/organizations/{orgId}/ai/reply
Authorization: Bearer YOUR_API_KEY
Content-Type: application/json
{ "ticketId": "uuid" }
# Summarize a ticket conversation
POST /v1/organizations/{orgId}/ai/summary
{ "ticketId": "uuid" }
# Translate a message
POST /v1/organizations/{orgId}/ai/translate
{ "text": "Bonjour", "targetLanguage": "en" }
# Get / update AI settings
GET /v1/organizations/{orgId}/ai/settings
PATCH /v1/organizations/{orgId}/ai/settings
{ "ai_assistant_enabled": true, "ai_response_style": "concise" }ⓘBilling
confidence score — build your own escalation logic based on it.AI model
Talki AI is powered by Google Gemini. The knowledge base uses vector embeddings for semantic search — the AI finds relevant content even when the customer's question doesn't use the exact same words as your documentation. No model configuration is required on your side.