One platform connecting document parsing, RAG retrieval, Deep Agent and LLM Wiki —
so knowledge answers accurately, delivers registers and reports, and keeps improving inside its permission boundary
Current capabilities grouped by workflow stage. Follow the linked docs for configuration detail.
| Category | Feature | What it does | Docs |
|---|---|---|---|
| Parsing and chunking8 features | Multi-format ingestion | Handles PDF, Word, Excel, PPT, Markdown, images and video, with parsing running asynchronously. | Read docs |
| MinerU / PaddleOCR | Parses scans, complex layouts, tables, formulas and images while keeping headings, page numbers and coordinates. | Read docs | |
| MinerU-Popo enhancement | Optionally strengthens heading hierarchy, section relations, cross-page tables and image context, falling back to plain MinerU on failure. | Read docs | |
| Five chunking strategies | Smart, Title, Regex, Parent-Child and Page chunking, chosen by document structure and retrieval goal. | Read docs | |
| Parent-child chunking | Match precisely on small chunks, then return the fuller parent context, with the relationship maintained over time. | Read docs | |
| Complex table handling | Preserves very long tables, merged cells, repeated headers, context rows and cross-page continuity. | Read docs | |
| Image and video understanding | Extracts images, captions and context; video supports ASR, keyframes, VLM descriptions, timestamped citations and retrieval. | Read docs | |
| Preview and re-run | Inspect chunks and their source position online, then re-run processing after adjusting parsing or chunking settings. | Read docs | |
| Retrieval and Q&A9 features | Hybrid keyword and vector retrieval | Combines lexical recall, vector recall, weighting signals and optional reranking for everyday questions. | Read docs |
| Synonyms and metadata filters | Extend business terms via a knowledge base dictionary, and narrow retrieval by metadata, knowledge base or a named document. | Read docs | |
| Retrieval testing and debugging | Inspect coarse ranking, reranking, keywords, similarity scores and which path produced each hit. | Read docs | |
| DeepRead close reading | Reads across the table of contents, relevant body text and full sections in several steps, producing cited answers. | Read docs | |
| ColPali visual fusion | Native Milvus multi-vector retrieval runs alongside text retrieval and is merged with RRF — ideal for charts, decks, scans and layout-heavy content. | Read docs | |
| SAG multi-hop retrieval | A Milvus-based multi-hop engine whose results are fused with native retrieval, for cross-document facts and relationships; the strategy is chosen per assistant. | Read docs | |
| RAPTOR / GraphRAG | Adds global themes, entity relations and long-document hierarchy through summary trees and knowledge graphs. | Read docs | |
| Traceable citations | Citations return to the page, coordinates and image position, with PDF preview and highlighting for verification. | Read docs | |
| Web search fallback | Optional Bocha web search supplements answers with public material when internal knowledge falls short. | Read docs | |
| Deep Agent and deliverables14 features | Office file delivery | Built-in Word, PowerPoint and Excel skills generate files in an isolated sandbox; they appear as cards in the chat and preview in place. | Read docs |
| Isolated agent sandbox | Each session gets an offline, isolated workspace; uploaded attachments are placed in it and cleaned up when the session is deleted. | Read docs | |
| Skill catalogue | Admins upload skill packages that install in a separate environment; skills declare versions and environment variables, and users @-mention them per turn. | Read docs | |
| Long-term memory | Once enabled by an admin, remembers user preferences and facts across sessions; inferred memories wait for confirmation, and users can edit or clear them. | Read docs | |
| Steering while running | Keep typing while the agent works: add to the current turn, queue for the next one, or send automatically when it finishes. | Read docs | |
| Thinking switch | One assistant-level switch decides whether Q&A, agent, writing and vision models think before answering. | Read docs | |
| In-place @ mentions | Capabilities, skills and documents are all @-mentioned in the chat box and apply to the current turn only. | Read docs | |
| @document scoping | Name the exact material for this turn from authorised knowledge bases; plain Q&A and advanced skills share the same boundary. | Read docs | |
| Cross-section close reading | Browses document structure, retrieves body text and reads context per node — for policy interpretation, technical verification and comparison. | Read docs | |
| Numbered clause analysis | Counts clauses, finds the highest number, and checks continuity, gaps and duplicates for Chinese numbered articles. | Read docs | |
| Structured data extraction | Extracts records for your fields across one or many complete documents, separating reliable, missing and ambiguous values. | Read docs | |
| Table and Excel delivery | Turns verified tabular records into a downloadable register, keeping each record’s document identity and source. | Read docs | |
| Human input and task resume | Pauses when a key definition cannot be settled, then resumes from the current stage once you answer, without redoing finished work. | Read docs | |
| Report and article writing | Writes reports, proposals, scripts or articles from named material, with section-by-section streaming, image citations, in-place editing and Markdown export. | Read docs | |
| LLM Wiki knowledge network5 features | Auto-generated wiki | Builds entity, concept and summary pages from knowledge base material, turning a document pile into a browsable network. | Read docs |
| Full-text search | Search wiki pages, business entities and domain concepts by keyword instead of asking one question at a time. | Read docs | |
| Cross-page links | Creates readable links between related pages so readers can follow concepts and dependencies. | Read docs | |
| Relationship graph | Visualises how entities and pages relate, keeping the underlying material and citations. | Read docs | |
| Incremental upkeep and health checks | Maintains affected pages as documents are added, removed or updated, with generation status, checks and recovery. | Read docs | |
| Knowledge management and operations8 features | Knowledge directory tree | Manage many knowledge bases in nested folders — create, move, rename and browse by directory. | Read docs |
| Import and export | Migrate knowledge bases between environments, deliver offline and govern in bulk, avoiding repeated parsing. | Read docs | |
| Unified file management | View, upload, download, move and manage files centrally, with visibility and actions bounded by permissions. | Read docs | |
| Operations overview | See usage by knowledge base, question, retrieval hit, user and time — filtered by the current user’s permissions. | Read docs | |
| Knowledge gaps and feedback | Aggregates misses, low-quality answers and user feedback into a concrete improvement backlog. | Read docs | |
| AI-assisted diagnosis | Analyses causes from answer feedback and retrieval evidence, one at a time or in bulk, and proposes fixes. | Read docs | |
| Tasks and operational reports | Track background task status and export operational detail and periodic reports for the current filter scope. | Read docs | |
| Per-user model configuration | Administrators set default models and permitted scope per user, validating authorisation and syncing credentials. | Read docs | |
| Organisation, permissions and identity8 features | Users and nested organisations | Maintain users, groups, departments and sub-organisations, with primary organisation, members and org admins. | Read docs |
| Collaboration groups | Create horizontal groups for cross-department projects, with their own members and knowledge base grants. | Read docs | |
| Pure RBAC authorisation | Grant view, edit or manage rights to users, organisations and collaboration groups — no tenant field standing in for real authorisation. | Read docs | |
| Directory inheritance | Grants on a parent folder flow down to sub-folders and knowledge bases, with the highest effective permission winning. | Read docs | |
| Org admin boundaries | Org admins manage only their authorised subtree — lists, retrieval, operations and exports all share that scope. | Read docs | |
| WeCom sign-in and org sync | Browser QR and WeCom workbench sign-in, with account mapping and organisation sync on first login. | Read docs | |
| External OAuth2 / OIDC | Integrates with enterprise identity providers and Keycloak OIDC, handling code exchange, sessions and profile mapping. | Read docs | |
| Shared conversation access control | Shared conversations can require sign-in and limit knowledge to the visitor’s own RBAC scope. | Read docs | |
| Integrations and open APIs9 features | WeCom bot | Brings knowledge Q&A into WeCom conversations, including internal work entry points and rich replies. | Read docs |
| DingTalk bot | Configure the DingTalk app, bot permissions and callbacks to use KnowFlow inside DingTalk. | Read docs | |
| Feishu bot | Connect knowledge base capability to Feishu conversations through app and bot permission configuration. | Read docs | |
| Dify integration | Connect via external knowledge base API, plugin or workflow, returning traceable citations. | Read docs | |
| MaxKB integration | Bring KnowFlow retrieval into MaxKB workflows through API tools and advanced orchestration. | Read docs | |
| REST API and SDK | Integrate internal systems, automation and your own agents through standard interfaces and SDKs. | Read docs | |
| Live API reference | Swagger UI at /api/docs and Redoc at /api/redoc are generated from code and testable with an API key; chat streams can return deltas only. | Read docs | |
| MCP access | List knowledge bases and assistants over MCP, with a ready-made connector for WorkBuddy (Tencent). | Read docs | |
| Sharing and embedding | Standalone share pages, full-screen embedding and web widget previews for existing portals. | Read docs | |
| Self-hosting and infrastructure8 features | Docker Compose | Complete container orchestration and environment configuration, with core services and optional capabilities enabled as needed. | Read docs |
| Kubernetes / Helm | Distributed task execution, replicas and service separation for enterprise clusters and high availability. | Read docs | |
| Metadata database options | Supports PostgreSQL and MySQL, with a migration tool from MySQL to PostgreSQL. | Read docs | |
| Document engine options | Supports Milvus, Elasticsearch and Infinity — choose by data size, retrieval style and operational constraints. | Read docs | |
| Object storage options | Supports MinIO and RustFS for self-hosted files, parsing artefacts and image assets. | Read docs | |
| Offline environments | Offline images, models and knowledge base delivery keep data and models entirely inside your network. | Read docs | |
| Optional services | SAG, ColPali and MinerU-Popo deploy separately, so the base environment does not carry every GPU and service cost. | Read docs | |
| Runtime diagnostics | Health checks, task queues and the actual running engine are exposed so administrators can locate deployment issues. | Read docs |
Some capabilities require a separate service, a specific model or an optional deployment component. Actual availability depends on the linked docs and your deployment configuration.
From knowledge entering the system, through retrieval and linking, to governance and continuous improvement
Understand the document before chopping up the text
KnowFlow keeps heading hierarchy, paragraph relations, tables, formulas, images and layout position, then picks the right knowledge unit per document structure — a traceable base for stable retrieval.
Route each question down a path that matches its difficulty
Everyday questions use mature keyword-plus-vector hybrid retrieval; complex ones go through SAG for cross-document, multi-step search; charts and layout-heavy content are covered by the ColPali visual path.
Not just answers — registers and finished documents
Deep Agent performs multi-step reading, field extraction and writing inside the user’s authorised scope, turning enterprise documents into verifiable, downloadable and editable work output.
From a pile of documents to an explorable network
LLM Wiki generates a structured encyclopedia from the knowledge base, organising entities, concepts, pages, relations and sources so people can ask, search, browse and follow relationships.
From "the documents are loaded" to "the knowledge keeps improving"
Real questions, retrieval hits and user feedback become observable metrics and knowledge gaps, then AI diagnosis, improvement tasks and reports drive quality up over time.
Keep knowledge structure aligned with how the company is run
Build multi-level organisations, members and collaboration groups, then organise many knowledge bases in a directory tree with clear ownership, business boundaries and maintenance responsibility.
The right knowledge reaches the right people
The super admin lays out top-level folders and assigns folder managers, who then create knowledge bases and grant access within their own folders. Any directory or knowledge base can be granted to users, organisations or collaboration groups; permissions inherit downward, the highest grant wins, and the same boundary applies to retrieval, management and operational data.
Parsing, retrieval, knowledge network, operations, governance and self-hosted infrastructure
How KnowFlow differs from conventional document Q&A
| Dimension | KnowFlow | Conventional approach |
|---|---|---|
| Knowledge unit | Keeps headings, parent-child hierarchy, tables, images and layout | Fixed-length splitting that loses structure and context |
| Retrieval paths | Hybrid retrieval, SAG multi-hop and ColPali visual retrieval working together | Every question shares one vector search path |
| Knowledge form | Beyond Q&A, a searchable and browsable LLM Wiki knowledge graph | Knowledge stays in a file list and a chat window |
| Ongoing operations | Metrics, knowledge gaps, AI diagnosis, improvement tasks and reports form a loop | No quality feedback or improvement handle after launch |
| Knowledge organisation | Organisations and a directory tree manage ownership and business boundaries | Flat knowledge bases that get unmanageable at scale |
| Permission governance | Directories or knowledge bases granted to users, organisations and groups, with inheritance | Usually single-base or simple member permissions |
| Deployment and data | On-premise and offline deployment, with data and models under your control | Depends on external cloud services with platform-defined boundaries |
From precise Q&A to complex research and group-wide governance
Sections, clauses, tables and context are preserved, so structure-aware RAG gives stable, accurate, traceable answers.
SAG performs cross-document retrieval and step-by-step verification for questions that combine facts or trace relationships.
LLM Wiki captures entities, concepts and relationships into a searchable, browsable and explorable knowledge graph.
Organisations, the directory tree and directory-level RBAC manage complex boundaries while the operations loop keeps quality improving.
Bring your documents, typical questions, organisation structure and permission rules, and evaluate the full loop with us