Intelligence · Module
Knowledge
Answers from your own documents, with the source, or an honest "I don't know"
Knowledge makes a company's documents, SOPs, policies and help content answerable by people, Copilot, Service Desk and Digital Employees. Every answer cites its source and respects source permissions, and unanswered questions are turned into content.
The retrieval foundation for One3
Knowledge is where a company's documents become answers. It is planned from W0 and is the base that Copilot, Service Desk suggested replies, the Help Center, Research and Digital Employees all rely on. A Knowledge Starter with limited collections and documents is included in the Platform; the Intelligence app removes those limits.
How content gets in
Content lives in collections, each with a visibility (restricted, tenant or public), a classification, a default access list and its own chunking profile. W0 covers uploads of PDF, DOCX, HTML and Markdown with OCR for scanned PDFs and tables kept in structured form, plus single-page URL ingestion. W2 is planned to add a website crawler that respects robots.txt and connectors for Google Drive, Gmail, Notion and websites with permission sync, so a file shared with one person stays visible to that person only. SharePoint, OneDrive, Confluence, Slack, Teams and Zoho WorkDrive are planned for W3 (API terms to be verified).
At ingest, documents are scanned for personal data, including Indian identifiers such as PAN, masked Aadhaar, GSTIN and IFSC. Restricted content found in a tenant-wide collection is quarantined. Unchanged documents are skipped by content hash.
How answers are made
Retrieval combines vector search and full-text search, then reranks the results. Only chunks the asker is allowed to see are considered; a leak test runs in CI. The answer cites document, version, page and quote span, passes a groundedness check and shows a freshness indicator. If confidence is too low, the user sees "I don't know" and search results instead. Compliance collections can be set to quote-only, so no paraphrase is generated.
Knowledge that maintains itself
Unanswered and refused questions are clustered weekly into gaps with example queries. Each gap can become a draft article in Content Hub. The gap only closes when the published article is indexed and the example questions now answer. Owners get review-by dates and a "verified by owner" badge, and the system flags duplicate and contradicting documents.
Governed for people and agents
Agents get knowledge scopes: their effective access is the granted collections intersected with their owner's permissions. Tools are exposed as knowledge.search and knowledge.ask over API and MCP. DPDP erasure is planned to reach documents, chunks, embeddings, answers and feedback (verify with counsel), and legal hold blocks deletion with a stated reason.
Features
What Knowledge includes
Collections with visibility and ACLs
Restricted, tenant or public collections, each with classification, default access list, embedding model and chunking profile.
Upload with OCR and table structure
PDF, DOCX, HTML and Markdown; OCR for scanned PDFs; tables kept in structured form for cell-level answers.
Connectors with permission sync
Google Drive, Gmail, Notion and website crawler planned for W2, mirroring source sharing; more connectors in W3 (verify API terms).
PII scan at ingest
Detects personal data including PAN, masked Aadhaar, GSTIN, IFSC and phone numbers; suggests classification and quarantines restricted content.
Hybrid retrieval with reranking
Vector and full-text candidates fused by reciprocal rank fusion, then reranked by a cross-encoder.
Permission-aware retrieval
Per-actor allowed access keys for users, groups, roles and agents, invalidated on any group or ACL change.
Citations and groundedness check
Every claim links to document, version, page and quote span; answers without citations are not shown.
Answer modes
Concise, step-by-step for SOPs, or quote-only for compliance collections where paraphrase is not allowed.
Multilingual questions
English at MVP; Hindi and Tamil questions over English and Indic documents planned at GA, evaluated per language.
Stale-answer invalidation
Answers citing a changed document are marked stale and re-run on demand.
Gap detection to draft article
Weekly clustering of no-answer queries into gaps that become Content Hub drafts for review.
Ownership and review dates
Document owners, review-by dates, expiry reminders and a "verified by owner" badge in citations.
Duplicate and conflict detection
Finds near-duplicate passages with contradicting statements and opens a resolve task.
Evals and knowledge analytics
Nightly recall, groundedness and citation accuracy per collection; answer rate, no-answer rate and cost per answer by surface.
Workflow
Ingestion and permission sync
- 1
A source is uploaded or connected
- 2
A sync job checks the content hash and skips unchanged documents
- 3
New versions are parsed, OCR'd, PII-scanned and classified
- 4
Restricted content in a tenant-visible collection is quarantined and the uploader notified
- 5
Clean content is chunked with heading paths and embedded
- 6
A source permission change recomputes access keys for every chunk in one transaction
Workflow
Gap to published content
- 1
No-answer and refused queries are clustered weekly
- 2
A gap with example queries and counts is created
- 3
A draft article is created in Content Hub
- 4
A person reviews and publishes it
- 5
The gap resolves once the example queries now answer
AI
AI skills and Digital Employees
Answer generation with citations
Read-only answers with citations, confidence and a groundedness check.
Query rewrite and decomposition
Splits multi-part questions into sub-queries before retrieval.
Gap clustering and draft articles
Groups unanswered questions and drafts articles; publishing is human.
SOP writer
Drafts a structured SOP from notes, transcripts or a process description, with required human review.
Prompt-injection heuristics
Flags chunks that look like instructions to the model at ingestion.
Digital Employees that use this module
They act through the same governed Actions as people; risky steps wait for approval.
Frequently asked questions
Is Knowledge included with the Platform?
A Knowledge Starter with limited collections and documents is included in the Platform. When limits are reached, new ingestion pauses with an upgrade prompt and existing content stays answerable.
Will people see documents they should not?
Retrieval only considers chunks the asker is permitted to see, and connector permissions are mirrored from the source. A permission leak test is part of CI and each connector's release gate.
Which languages are supported?
English at MVP. Hindi and Tamil questions are planned at GA, enabled per language after evaluation.
Can agents use Knowledge?
Yes, through knowledge scopes. An agent sees only the collections granted to it, and never more than its human owner.
See Knowledge in your workflow
Tell us about your team and the outcomes you want. We will walk you through the apps, the Digital Employees and the plan for your industry.