It has to read
Headings, tables, images and layout are preserved so retrieval works on the original meaning of a document.
KnowFlow · Enterprise knowledge and data, built to be verified
Every answerholds up to scrutiny
Complex documents are understood first, then answered. Parsing, knowledge Q&A, Deep Agent and traceable delivery all run under directory-level RBAC. Deploy on-premise or fully offline so knowledge stays usable, permissions stay enforced and data never leaves your network.
- Structure first, answers trace back to the sourceHeading hierarchy, tables and layout are preserved, so answers point to a page and a passage.
- Augmented retrieval across documentsSAG handles multi-hop questions; LLM Wiki connects entities, concepts and sources.
- Execution that produces real deliverablesDeep Agent reads deeply, extracts fields and writes reports or spreadsheets — every conclusion keeps its citation.
Current user's accessible scope Corporate network
Organisation
Directory permissions
Directory permissions
▾ Group HQ ▾ R&D centre ▰ Product team ▱ Project A ▱ Compliance & risk ▱ Finance
Multi-path retrievalStructuralSemanticVisual
Knowledge Q&A / Deep AgentAnswer directly or deliver a report
Evidence E1 · E2 · E3
Within authorised knowledge onlyGoverned delivery▤ Structured data▧ Analysis report▣ Business brief
Three non-negotiables for an enterprise knowledge system
It has to be governed
Directory-level RBAC runs through Q&A, Deep Agent and knowledge operations — not just the admin console.
It has to stay in-house
On-premise and offline deployment keep data, models and components under your control.
TWO PRODUCT LINES
Documents and databases are different problems
Every company has documents nobody can finish reading and databases nobody can query. KnowFlow solves both with two product lines that share one account system, one permission model and one deployment story.
Knowledge Base
Structure-aware understanding turns PDFs, Word files, scans, images and video into knowledge you can search, trace and deliver on.
- Deep parsing with five chunking strategies
- Structural, semantic and visual retrieval paths
- Deep Agent for close reading, extraction and reports
- Directory-level RBAC across retrieval and operations
Analytics
Metrics, dimensions and definitions become a governed semantic layer. Colleagues ask in plain language; a deterministic compiler writes the physical SQL.
- Governed semantic catalog and business dictionary
- The LLM writes S2SQL; the compiler writes SQL
- Read-only guard with fail-closed execution
- Fixed-stage diagnostics and vocabulary feedback loops
01
STRUCTURE-AWARE RAG
Understand the structure first, then talk about retrieval quality
Meaning in enterprise material lives in heading hierarchy, table relationships, page layout and context. KnowFlow keeps that structure so every piece of evidence can be located and checked.
DIRECTORY-LEVEL RBAC
Least-privilege access that still supports collaboration and compliance
Permissions flow from your organisation into the knowledge directory, then automatically through retrieval, agents and operational data. What buyers get is a governance chain, not a permissions feature.
Organisation and grantees
- Users
- Multi-level organisations
- Cross-team collaboration groups
Delegated directories
- Super admin creates top-level folders
- Folder managers are assigned
- They create knowledge bases inside
Inherited directory permissions
- Read-only / edit / manage
- Inherited from parent folders
- Highest grant wins
One boundary everywhere
- Retrieval and Q&A
- Deep Agent
- Operational data and exports
Permissions are not a separate module. The same user only ever touches their accessible directories — in Q&A, in Deep Agent, in operations and in exports.
PRIVATE & OFFLINE DEPLOYMENT
Self-hosted and offline, with data and models under your control
Fits corporate networks, air-gapped environments and existing infrastructure. Applications, services, models, storage and permission boundaries all sit inside your own security perimeter.
Application layer
Employee portalWeb / APIOA / ERP / PLMCollaboration tools
Service layer
Retrieval serviceDeep AgentDocument parsingPermissions & audit
Intelligence layer
Structural retrievalSemantic retrievalVisual retrievalModel serving
Infrastructure layer
CPU / GPUObject & block storageInternal networkBackup & DR
On-premisePostgreSQLVector databaseObject storageBackup & DR
Optional servicesStructure parsingVision modelsLarge language modelsOCR / layout
TLS in transitAES-256 at restRBAC access controlOperation audit logBackup & disaster recovery
MULTI-PATH RETRIEVAL
Complementary retrieval paths lift both recall and precision
Ordinary RAG pushes every question down one path, so complex documents lose hierarchy, tables and layout. KnowFlow makes the paths work together, then returns to the source evidence.
- Broader coverage: complementary paths reduce misses
- Steadier reasoning: narrow the scope before generating
- More control: paths, evidence and results are traceable
Ordinary RAGUser questionVector searchTop-K passagesGenerate
KnowFlow · fused pathsStructuralDirectory / heading / level SemanticVector / keyword VisualPage / chart / layout
Narrow by document structure first
Covers stable, high-frequency questions
Adds chart and layout evidence
Fuse and rerank → high-quality context → answer with evidence
CORE KNOWLEDGE Q&A
Every employee gets a verifiable answer, inside their permissions
For everyday questions, employees pick a knowledge base or a specific document and ask. The system retrieves with conversation context, cites its sources in the answer and supports follow-ups.
Pick one or several knowledge bases@document to scope this turn preciselyCitations return to the page and passageMulti-turn sessions keep the context
Available through WeCom, DingTalk, Feishu, Dify and internal systems, so answers reach employees where they already work.
What approvals are needed for travel expenses above ¥5,000?
K
Per the Expense Reimbursement Policy, three levels of approval are required:
- The direct manager confirms the expense is genuine;
- The department head reviews budget and business need;
- The finance lead completes the compliance review.
Single claims above ¥20,000 also require the responsible executive.
Follow-up: what if the traveller and the approver are the same person? →
DEEP AGENT
Deliver auditable, traceable business output
Deep Agent reads deeply within the current user’s authorised scope, performs structured extraction and analysis, writes the report, and links every conclusion back to its evidence.
REAL WORKFLOWS
Put complex knowledge work inside real business processes
Financial analysis
Pull metrics across reports, verify definitions and produce a sourced working paper.
Legal and compliance
Locate clauses and compare differences within authorised contracts and policies.
Human resources
Policy answers follow the employee’s organisation, role and directory permissions.
Engineering
Connect standards, design material and project documents while keeping the evidence chain.
Supply chain
Aggregate supplier, quality and delivery material to support complex investigations.
POC WITH YOUR DATA
Validate with your own material
Do not decide on generic demo data. Bring representative documents, real permission structures and real business questions, and settle it in one POC.
- Are complex documents parsed correctly?
- Do key questions find complete evidence?
- Do permission boundaries hold across the product?
- Does deployment fit your environment?