Skip to main content
Features

The full job description.

Every capability that makes ZenSearch a coworker rather than a chatbot — answers that run agents, actions that wait for approval, automations that work unattended, and the retrieval engine underneath it all.

Agentic Answers

Every question runs an agent

Ask a question and ZenSearch runs a budget-tiered agent loop over your connected sources — not a single canned search. It reasons across your knowledge base, databases, the web, and live connector reads, and cross-checks them instead of trusting the first hit.

  • Three answer modes per message: Chat (knowledge base only), Auto (the agentic default), and Research (full planning with approval-gated write tools)
  • Cross-source corroboration — a Jira ticket and its linked GitHub PR get reconciled rather than guessed
  • Transparent by default: inline reasoning, tool retries, delegation hand-offs, and a live cost meter while the agent works
  • Verified citations and a confidence score on every claim
Demo / Agentic Answersagentic-answers
ChatAutoResearch
Did we ship ZEN-213?
▸ search_jira_issues → ZEN-213: In Review
▸ search_github → PR #1188 merged
corroborating sources…
✓ Jira says In Review, but the linked PR merged — likely Done
3 sources · 2 tools$0.018

Approval-Gated Actions

Autonomy with a human in the loop

The coworker doesn't just read — it posts replies, creates tickets, updates records, and sends email. Every write carries a risk tier, and the consequential ones pause the run until the right human says go.

  • Risk tiers T0–T5 decide what runs freely and what waits — with per-team policy overrides
  • Approvals route to the right person (preferred approver, on-call, or manager from your org chart) and are decided from Slack, an email reply, or the web
  • Drafts in your voice from your own account — or with explicit AI disclosure where team policy requires it
  • Every action attributed and audit-logged with the full reasoning trace
Demo / Approval-Gated Actionsactions-approvals
update_zendesk_ticket — reply drafted for #4821
⏸ T4 — approval required→ Maya R. (on-call)
✓ approved via Slack · 11:47
✓ reply posted with citations
voice: yours — sent from your accountaudit: full trace

Self-Improving Agents

Better in month two than on day one

Enterprise agents that learn from successful workflows, delegate to specialists, and synthesize complex answers with complete audit trails.

  • 100+ pre-built tools — search, SQL, Google Workspace, M365, Zendesk, ServiceNow, HubSpot, Salesforce, GitHub, and more — that read and take approval-gated actions
  • Extensible with custom webhook tools, MCP servers (SSE, streamable HTTP, stdio), and sandboxed code execution
  • Multi-agent delegation: subtasks hand off to specialist agents (depth-limited), including approval-gated writes
  • Three-layer memory: persistent facts, observational compression (80%+ token savings), and procedural memory that learns reusable workflows
  • Dynamic budget tiers + pre-flight cost ceiling: every run is capped in dollars before the first token is spent
  • Pause & resume on budget exhaustion — partial answer plus a Continue button, with full state preserved for 7 days
Demo / Self-Improving Agentsai-agents
Agent: research-assistant
▸ init_node
▸ plan_node → 3 steps
▸ tool: search_documents
found 12 results
▸ tool: query_database
SELECT * FROM users...
▸ synthesize_node
✓ complete (1.2s)

Automations & Triggers

Works while you don't

An automation is an agent with a job description and a trigger. Schedule it, email it, or wire it to your systems — it runs in the same governed runtime as every chat.

  • Cron schedules with verification contracts and bounded auto-resume on paused runs
  • Inbound email: every automation can have its own address — replies to its messages resume the run
  • Event webhooks from Jira, Zendesk, GitHub, Confluence, Salesforce, and HubSpot, attributed to the triggering user
  • Meeting events from Zoom, Google Meet, and Microsoft Teams; results delivered via webhook, Slack, or email
Demo / Automations & Triggersunattended
Triggers
cron Mon 07:00 — leadership brief
webhook Zendesk · ticket.created
meeting Zoom · recording.completed
4 automations activeauto-resume on pause

Advanced Search

Precision meets intelligence

Combines keyword precision with vector search and AI-powered semantic understanding. Accurate results whether people search with exact terms or natural-language questions.

  • Semantic and keyword search fusion for optimal relevance
  • AI-powered reranking ensures top results are always accurate
  • Automatic query enhancement for improved recall
  • Faceted filtering across departments, topics, dates, and sources
Demo / Advanced Searchhybrid-search
authentication flow
Dense similarity0.89
Sparse BM250.76
Final (reranked)0.94

Natural Language to SQL

No SQL required

Connect PostgreSQL, MySQL, ClickHouse, or SQL Server. Ask questions in plain English — ZenSearch discovers the schema and writes the query.

  • Automatic schema discovery and relationship mapping
  • Read-only execution with query validation
  • Error correction with LLM-friendly feedback
  • SSRF protection for secure database access
Demo / Natural Language to SQLdatabase-search
Query: top customers by revenue last month
Generated SQL:
SELECT c.name, SUM(o.total)
FROM customers c
JOIN orders o ON c.id = o.customer_id
WHERE o.created_at > NOW() - '1 month'
GROUP BY c.id ORDER BY 2 DESC
LIMIT 10
Query validated — read-only

Verified Citations

Source transparency

Every answer includes traceable references to source documents — PDF pages, presentation slides, and exact code locations, with relevance scores.

  • Precise page references for PDF documents
  • Slide-level citations for presentations
  • File and line number references for code
  • Relevance confidence scores for each source
Demo / Verified Citationscitations
The rate limit is set to 1000 requests per minute...
api-docs.pdf:42config.yaml:15
95% relevanceVerified

Enterprise Security Controls

Multi-layer protection

Validates inputs, monitors outputs, and enforces security policies at every layer. Configurable per team with real-time policy updates.

  • Detects and blocks adversarial prompt-injection attempts
  • Identifies sensitive information including SSNs and credit cards
  • Verifies AI responses against source documents for accuracy
  • Team-level configuration with instant policy propagation
Demo / Enterprise Security Controlsguardrails
✓ Content moderationPASS
✓ Injection detectionPASS
⚠ PII detectedREDACTED
✓ Hallucination check0.92 confidence

Adaptive Response Formatting

Context-aware answers

Automatically detects query intent and adapts the retrieval strategy. Lists for enumeration, steps for procedures, comparisons for alternatives.

  • Comprehensive coverage for list-based queries
  • Hierarchical context for architectural questions
  • Sequential ordering for step-by-step procedures
  • Side-by-side analysis for comparison queries
Demo / Adaptive Response Formattinganswer-shapes
enumerative
list all APIs
procedural
how to deploy
comparative
X vs Y
exploratory
what is...
→ Retrieval strategy adapts automatically

Go Deeper

Three places to keep reading.

The Platform

Context, action, governance, and memory — the runtime that turns 18 sources and 100+ tools into a coworker.

Platform overview

Governance & Admin

Risk tiers T0–T5, cross-surface approvals, identity federation, memory policies, and the Control Tower dashboard.

Governance details

Every Surface

Slack, Teams, browser side panel, and web — the same coworker, with cross-surface continuity.

See surfaces

Under the Hood

And that's just the beginning.

A complete platform for an enterprise AI coworker. No assembly required.

18 Data Connectors

Native integrations for Confluence, Slack, GitHub, Notion, S3, SharePoint, Jira, Google Drive, and more.

Automated Pipeline

Event-driven ingestion that keeps the knowledge base fresh — smart dedup, scheduled syncs, and automatic cleanup of removed content.

Granular Access Control

Team-level roles and document-level permissions that sync with your source systems.

Centralized Model Gateway

Route all AI model calls through a unified proxy with usage tracking and rate limiting — swap providers without touching agents.

Org Awareness

Org chart, on-call schedules, and service ownership inform approvals, escalations, and paging.

Control Tower

Admin dashboard for connectors, costs, agents, automations, failed runs, and sync monitoring.

Provider Fallback & Circuit Breakers

Automatic failover between LLM providers — unhealthy providers are skipped until they recover, so one upstream incident can't take the coworker offline.

OpenTelemetry Tracing

End-to-end distributed tracing across services, model calls, and agent tool invocations. Export to Tempo, Jaeger, Honeycomb, or Datadog.

Get Started

See it
for yourself.

Every feature on this page is live in the demo — or talk to our team about a deployment shaped to your environment.