Governed AI for Critical Knowledge
Award-winning enterprise AI with complete accountability. Every answer cited, every action audited, every change traced. Run fully offline with open source models. Enterprise governance built in, not bolted on.
From Documents to Intelligent Answers
Upload Your Knowledge
Import 25+ document formats. iKB processes, indexes, and builds knowledge graphs automatically.
AI Understands Context
Advanced RAG pipeline with intent classification, contextual chunking, and entity extraction.
Users Ask Questions
Across 6 channels — or in Freeform mode with personal uploads and web search.
Accurate, Traced Answers
Quality-scored, cited responses with full pipeline provenance tracing.
Personal AI Workspace
A new conversation mode where users chat with AI without a pre-configured knowledge base. Upload personal documents, search the web, and get answers — all within governed limits.
Standard KB chat and Freeform mode, switchable via sidebar toolbar
Searches both uploaded personal documents and web simultaneously
Comprehensive coverage for quotas, model validation, conversation limits
Full CRUD endpoints for conversations, uploads, and quota checking
Core Capabilities
Users upload documents into their own Freeform conversation with web search fallback when local docs are insufficient.
Administration
Admins control Freeform access, quotas, and model availability with full session visibility.
Per-User Semantic Memory
AI remembers user preferences and facts across sessions. Powered by Mem0 with pgvector. Hybrid architecture: global memories shared across topics + topic-specific memories scoped per domain.
Automatic Extraction
After each conversation turn, the AI silently learns relevant facts about the user — preferences, role, style, and domain knowledge.
Privacy-First
Incognito mode skips memory entirely. Full GDPR/PDPA DSAR support: search, export, and purge user memories. Every operation audit-logged.
All 6 Channels
Memory works across Native, WebChat, Slack, Teams, OMNI, and API. Anonymous users get session-scoped memory with auto-cleanup.
Unified AI Model Gateway
Full LiteLLM integration as the unified model gateway, replacing direct API calls. Provider-first model discovery, per-model hybrid routing, unified cost tracking, and dynamic configuration — all managed from the admin UI.
Provider Discovery
Discover and enable models from OpenAI, Anthropic, Google, Cohere, Cerebras, and LocalAI/Custom providers directly from admin UI.
Hybrid Routing
Each model independently routed through LiteLLM or legacy direct API. "iKB LLM Router" / "Legacy" badges in all dropdowns.
Cost Tracking
Real-time cost dashboard with per-topic, per-model, per-channel spend attribution via LiteLLM's extra_spend_tag_headers.
18 Built-In Output Skills
A complete output format templating system that controls how AI structures responses. 18 system skills covering retrieval, professional documents, and export formats — plus a custom skill builder with governance controls.
Direct Answer, Summary, Step-by-Step, Comparison, Checklist, FAQ, Briefing Note, Email Draft, Report, Proposal, Word/Excel/PPT exports
Manual @prefix (free) → regex patterns (free) → intent mapping (free) → LLM classifier (tokens)
Forbidden prompt injection patterns validated on save. Skill instructions explicitly lower-priority than system rules.
Manual @prefix on all channels. Auto-detection on native chat. Pipeline tracing logs skill detection method.
Custom Skill Builder
Create custom skills with name, slug, prompt template (2000 chars), regex triggers with ReDoS validation, and governance flags.
Document Export
Per-message export to Word, Excel, and PowerPoint via ⋮ menu. AI-powered reformatting with Advanced Processing Model. Auto-export on narrative requests. Diagrams embedded in exports.
Information, Terms & Privacy
Comprehensive consent management with governance-first design. Blocking modal on first visit, admin-facing responsible administration interstitial, and WebChat widget terms — all with immutable audit trails and version tracking.
User & Admin Terms
Blocking acceptance modal for users. "Responsible Administration" interstitial for admins. Decline = redirect/logout. Version bumps force re-acceptance.
Governance
Immutable TermsAcceptance audit trail. CORS-enabled widget endpoints. Rate-limited (30/min status, 5/min accept). Per-language content via AppSettings.
Automated Quality Testing
Evaluate AI response quality using the promptfoo testing framework. Create test cases, run evaluations, and measure accuracy at scale.
promptfoo Integration
Run evaluations against topics using the promptfoo binary. Start, monitor, cancel, and purge evaluation runs with real-time polling.
Bulk Test Import
Import test cases from CSV/XLSX with header normalization, BOM handling, delimiter detection, and Unicode support.
Results Dashboard
View pass/fail rates, assertion details, and response quality metrics. Automated scoring for empty, error, short, and unhelpful responses.
Per-Admin DACL Permissions
18 granular permission columns per admin account with deny-by-default enforcement. Topic-scoped and system-scoped tiers with hierarchy enforcement to prevent privilege escalation.
Topic-scoped: model selection, web crawl, token costs, documents, analytics. System-scoped: users, channels, settings, models, governance
Topic-scoped permissions and System-scoped permissions — deny-by-default on both
Every permission check enforced at API level, not just UI. Fail-closed on missing permissions.
All permission changes tracked in governance audit trail with actor and timestamp
Admin Management
Superadmins assign granular permissions per admin. Visual permission matrix with bulk operations. Hierarchy enforcement prevents privilege escalation.
Enforcement
Server-side middleware checks permissions on every API call. UI dynamically hides unauthorized sections. Per-channel-type creation limits (max Slack bots, max Teams channels).
7 Visualization Engines
AI responses come alive with interactive charts, mind maps, maps, timelines, inline SVG graphics, dashboard grid layouts, and syntax-highlighted code — all rendered client-side with lazy loading.
ECharts
Interactive charts and data visualizations. Bar, line, pie, scatter, radar — AI generates chart configs from data analysis.
Markmap
Mind map rendering from markdown headings. Interactive zoom, pan, and collapse for complex knowledge structures.
Leaflet
Interactive maps with markers, popups, and tile layers. AI can plot locations, routes, and geographic data.
Prism.js Syntax Highlighting
Beautiful code blocks with language-specific syntax highlighting. 40+ language support with one-click copy.
vis-timeline
Interactive timeline visualizations for historical events, project milestones, and chronological data. Zoom, pan, and grouping.
Inline SVG Graphics
AI generates custom vector graphics directly in responses — architecture diagrams, schematics, infographics, comparisons. DOMPurify-sanitized, hidden until rendered.
Dashboard Grid Layouts
AI arranges 2–4 charts side by side using layout markers. Responsive — columns stack vertically on mobile.
Diagram Engine — Unified Tooling
Per-diagram toolbar with copy, download, and expand. All diagrams render in light mode regardless of page theme. AI-generated disclaimer on all containers. 70% transparent backgrounds. Native SVG embedding in zoom modal. Diagrams embedded in Office exports (DOCX/XLSX/PPTX).
Single Shared Pipeline — All 6 Channels
All channels share a single retrieval pipeline, eliminating duplicated code. Always-on tracing with governance-grade provenance. 8-stage pipeline with query decomposition and neighbourhood expansion.
8-Stage Retrieval Pipeline
Unified pipeline with per-document search weights, query decomposition for complex multi-part questions, and neighbourhood expansion for chunks that span boundaries. All channels gained these features automatically.
Technical Details Modal
Per-message pipeline trace visible in session detail admin page. Full LLM generation config stored per message (model, temperature, max_tokens, reasoning_effort).
Intent Classification
AI classifies queries into 7 intent types (factual, comparison, summarization, multi-hop, procedural, clarification, out-of-scope) to dynamically adjust retrieval strategy.
Advanced Retrieval & Quality
Unified 8-stage retrieval pipeline with quality scoring, per-user memory, per-document search weights, and first-person AI voice.
AI Judge & Quality Scoring
QAG faithfulness decomposition, few-shot calibration (65% → 77.5% consistency), position-weighted scoring with relevance tier badges. Abstention-aware — honest refusals score as PASS.
GraphRAG
Cross-document entity relationships, hybrid graph+vector queries. Retry button for failed documents, graceful partial failure, PostgreSQL lock safety.
Neighbourhood Expansion
Adjacent chunks auto-pulled to capture cross-boundary information. Smart dedup, reading-order interleaving, configurable window.
Enterprise-Grade AI Governance
Four independent, immutable logging systems. Consent management with version tracking. OpenTelemetry instrumentation. Every action captured. Nothing deleted.
Logs, Prompt Versions, Config Changes, Moderation Events
Before/after snapshots with IP attribution
Fail-closed, per-topic, per-channel
Append-only. Deletion returns 403.
Content Moderation
13-category moderation using OpenAI's omni-moderation model, free of charge. Configurable per topic, enforced across all channels.
DSAR Compliance
Full GDPR/PDPA compliance integrated into the Governance dashboard. Multi-channel user data discovery, export, and erasure.
Complete Sovereignty — Run Fully Offline
Deploy on your own GPU infrastructure using open source AI models through vLLM, Ollama, or any OpenAI-compatible endpoint. Every component runs locally. No data ever leaves your network. Zero external dependencies for defence, government, finance, and any environment where data must never cross the perimeter.
Enterprise Security
147-finding security audit across every admin page and channel. Egress policy system for AI tool execution. SQL injection, SSRF, XSS, DACL enforcement, credential leak prevention, and nonce-based tool delimiters.
Dual-Key Encryption
AES-256-GCM with user key + admin key. All secrets Fernet-encrypted. Auto-redacted in logs.
Enterprise SSO
Authentik OIDC, JIT provisioning, group sync. Three modes: SSO-only, hybrid, local-only.
App Hardening
SSRF blocking, ODBC injection prevention, CSP nonces, CSRF, HSTS, XSS encoding.
Account Lockout
5 failed logins triggers 15-minute lockout. Stored in DB, survives restarts. Auto-expiry.
Egress Policy System
Network egress control for AI tools. Per-tool allow/deny rules, DNS-time SSRF validation, nonce-based tool delimiters.
147-Finding Audit
Complete security audit: SQL injection, SSRF, XSS, DACL bypass, credential leaks, privilege escalation — all remediated.
Self-Hosted
Your infrastructure, your network. Air-gapped deployment. No data leaves the perimeter.
No External Training
Documents are NEVER used to train AI models. Complete data sovereignty guaranteed.
Session Security
HTTPOnly cookies, strict SameSite, UUID v4, TOTP 2FA, HMAC webhooks, IP whitelisting.
Fail2Ban Integration
Network-level brute force protection. Structured log format for Fail2Ban parsing. Auto-ban repeat offenders at firewall level.
XSS Remediation
Comprehensive cross-site scripting audit and remediation across all user-facing templates and API responses.
One Knowledge Base, Every Channel
Deploy across 6 channels with full cross-channel feature parity. Reasoning level and text verbosity now configurable on all channels. All RAG features and skills available everywhere.
Full Cross-Channel Feature Parity
All advanced RAG features — pipeline tracing, intent classification, content moderation, HyDE, multi-query, agentic retrieval, self-critique, reranking, GraphRAG, quality scoring, and negative response filtering — now work identically across all 6 channels. Unified governance prompt injection with smart layer truncation.
Web Chat Widget
Deploy on any website with a single script tag. Frosted glass input, logo in history sidebar.
External API Tools
MCP + REST tool calling with parallel execution, encrypted auth, confirmation gating, and execution audit.
Human Escalation & Chat Export
Human handover with email transcripts and rate limiting. Export conversations to Word, Excel, or PDF.
SQL Functions & Cloud Sources
Schema-level discovery, admin annotations, query playground, and 40+ cloud storage integrations via rclone. All 12 provider configs aligned with actual rclone options.
Powerful Admin, Delightful Chat
LiteLLM gateway management, skills builder, self-update, AI Firewall, instant tooltips, UI standardization, and consent management — all from the admin panel.
Admin Enhancements
Chat UX Improvements
25+ Formats, Smart Crawling
Document Formats
Magic byte validation, ClamAV scanning, ZIP bomb protection, UUID filename sanitization.
Web Crawling Enhancements
- Cloudflare Browser Rendering — Alternative crawl engine using Cloudflare's Browser Rendering API. Configurable per job alongside Playwright.
- Opt-In Document Download — Choose which document types to download during crawl (PDF, DOCX, etc.)
- Conditional Re-Crawl — HTTP conditional headers (ETag, Last-Modified) to skip unchanged pages
- Re-Crawl Completed Jobs — Not just failed ones. Force reprocess on retry for stale chunks.
Flexible Deployment, Any Scale
Cloud SaaS
Fully managed, automatic updates, Celery task hardening
Private Cloud
Dedicated instance in your cloud region
On-Premise
Your infrastructure, air-gapped available
Performance
Test Infrastructure
Covers: access control, account lockout, analytics, app settings, auth flow, celery tasks, circuit breaker, error handlers, eval service, freeform service, quality scoring, security, token counting, web crawl.
See iKB in Action
See how iKB can help make your documents more accessible and searchable.