Grounded, cited answers
Responses connect to approved sources and page references, with faithfulness scoring and grounding checks.
Search company knowledge, analyse data and prepare documents. Manage access, model selection and audit evidence in one platform.
Bring your sources together without losing their structure, language, or provenance.
Responses connect to approved sources and page references, with faithfulness scoring and grounding checks.
25+ document formats and 40+ storage connectors, plus databases, websites, APIs, and business systems.
Tool-selecting retrieval with query decomposition, hybrid search, reranking, and a grounding verifier.
An exact-phrase floor preserves verbatim clauses and named sections rather than dropping them to meet a retrieval budget.
GraphRAG combines graph and vector evidence in one ranked pool, with provenance on extracted facts, drift detection, and verified deletion.
Bilingual ingestion and structural anchors support Malay, Chinese, and Tamil content.
An automatically built, bilingual wiki for each knowledge domain, under administrator control.
Up to ten files per question, with retrieval and analytics support.
Analyse data, research a question, and prepare deliverables within the governed platform.
Workbook questions become validated SQL in a sandboxed, per-request DuckDB database. Results are independently checked rather than calculated by the language model. PDF table analytics are supported.
The analytics engine prevents blind aggregation across different currencies or units.
Self-hosted web search produces cited, multi-page briefings scoped to the conversation, with per-page citations, regional source prioritisation, and a pipeline audit trail.
Generate Word, PowerPoint, Excel, and PDF with sandboxed rendering, embedded knowledge-base images, and audited background exports that continue after a tab closes.
Sandboxed in-chat mini apps operate behind governed toggles and a global kill switch.
Charts, diagrams, infographics, maps, timelines, and equations include zoom and download controls.
Text-to-speech begins as output becomes available, with Malaysian Malay voice behaviour.
Private workspaces include user memory, token budgets, and DLP controls. Deliver through web chat, Slack, Teams, WhatsApp, Telegram, or REST API.
Control what the platform can access, how it operates, and what evidence it retains.
System architecture132 model providers, including xAI Grok, and local OpenAI-compatible endpoints. Set per-topic model policies and review Anthropic prompt caching and cost reporting.
Versioned prompts, access boundaries, moderation, injection controls, output scrubbing, and permission-gated grounded inference. Govern external tools and outbound destinations.
Users see retrieval stages. Administrators can inspect a root-cause pipeline analyser, stage timing, and outcomes.
Append-only, hash-chained events record prompt and configuration changes, sources, control decisions, and per-call model, token, and cost telemetry.
PII text and image redaction, DLP modes, DSAR workflows, and two-stage retention with physical purge.
Conversation evidence bundles use signed, expiring download links and notifications, with SIEM-ready audit records.
Preserve citations, provenance, control checks, and timing without disclosing the underlying model's hidden reasoning.
Hosted, private cloud, on-premise, and air-gapped options. Local storage, indexing, retrieval, models, caching, and background processing; controlled egress; a one-command operator CLI and 18-signal health probes.
Start with a bounded corpus and representative questions. Inspect retrieval quality, citations, failure cases, and the operating controls before rollout.
These are reported product evaluation figures, not a guarantee of accuracy on your data. Ask for the test corpus, methodology, and limitations during your technical review.
Discuss an evaluationBring representative documents and questions. Review the answers, source evidence, and controls with your team.