Answers cannot be verified
One wrong amount, deadline or approval rule is an incident. Without sources nobody dares to rely on it.
KNOWWORK · AI ASSISTANT FOR EVERY EMPLOYEE
Bring LLMs to work,
with sources and results
Employees ask in the browser or chat apps; the assistant searches company knowledge within their permissions, reads attachments and produces files.
Every citation opens the original text. Deployed on your servers, data never leaves.
Self-hosted
Cited answers
Department & folder permissions
WeCom / Feishu / DingTalk
Plugging in an LLM is easy
What comes after it enters the company is not
Handing employees a generic chat tool, or stitching together "upload docs + Q&A", demos well. Once rolled out, the problems lie outside the model.
Permissions leak
Put finance, HR and project files in one base and anyone can ask them out; split them into dozens and nobody maintains them.
It chats but does not work
Employees need "check these invoices against the new policy", not a paragraph they still have to rework.
Nobody can see the usage
Who asked what, which questions went unanswered, what it cost — after launch it is a black box.
KnowWork’s trade-off: the model understands and writes, permissions decide what it can read, commands run in an isolated sandbox, and every step is recorded.
How a question is handled
From the first message to filling knowledge gaps, every step is visible
- 01AskWeb / chat apps
Chat in the browser, or message / @ the bot in WeCom, Feishu, DingTalk
- 02SearchBy permission
Only the bases this person is granted, hybrid semantic + keyword recall
- 03ReadRead fully
When a hit is partial, read the surrounding text or section before answering
- 04ActSandbox + approval
Attachments, skills and files run in a session sandbox; writes need approval
- 05AnswerVerifiable
Each citation opens the original paragraph
- 06ReviewKnowledge gaps
Each turn is judged; unanswered questions go to the base maintainers
Each conversation gets its own sandbox container: commands and files stay inside and cannot touch the server, the database or other people’s data.
Six core capabilities
From managing knowledge to daily use and continuous improvement, in one system
01
Built-in knowledge base
Parsed on upload, searchable in the open
PDF, Word, PPT and Excel are parsed, chunked and indexed automatically; bases live in folders that carry permissions. The retrieval test page shows exactly what each question hits.
ParsingMinerU layout parsing, OCR, image captions, block-level manual fixes
Hybrid searchSemantic + keyword recall, reranking, synonyms, merged neighbours
Metadata & tagsCustom fields, auto extraction, filter by field and tag
VersionsDetects old and new versions of the same policy; answers follow the newer one
Maintainers can see what every document became and how it is searched.
02
Cited answers
Every number opens the source
The assistant searches, keeps reading when a passage is incomplete, then answers. Each citation shows the excerpt and opens the original file at that paragraph.
Precise locateJumps to the paragraph and highlights it in the preview
Figures inlineCited charts are placed back into the answer
Removable filtersMetadata filters are shown above the answer and can be removed
Quick answerKnowledge-only mode for fast one-shot questions
People trust it because every sentence leads back to the original.
03
Gets work done
Attachments, files and tools
Review a contract against policy, turn conclusions into Word, Excel or PPT, package company procedures as skills and connect business systems as tools.
Session sandboxOne isolated container per conversation, can be fully offline
Office outputsCreate and edit Word, Excel and PPT, preview and download
SkillsZIP packages with drafts, publishing and per-version success rates
MCP tools & approvalWrite and high-risk calls need the asker’s approval
Employees get results they can use directly.
04
Organisation & permissions
Who can read what, enforced
Workspaces isolate departments; within a space, grant folders, skills and tools to departments, groups or people, effective immediately. Company-wide bases are shared read-only instead of copied.
WorkspacesOne per department or subsidiary, each managing its own members and grants
Folder grantsUse / edit / manage; department grants flow down the tree
Cross-space sharingShare a base read-only to chosen or all spaces, revocable
SSOOIDC login; personal API keys never exceed their owner
Knowledge is organised by who may see it, so it never leaks across teams.
05
Ask in chat apps
WeCom, Feishu, DingTalk
Message the bot or @ it in a group; answers stream in and cited images follow. Long-lived connections — no public callback URL, no inbound ports.
All three platformsWeCom smart bot, Feishu custom app, DingTalk Stream mode
StreamingStreaming messages in WeCom, streaming cards in Feishu and DingTalk
ScopedEach bot only searches the bases chosen for it
Groups & DMsOne conversation per person and per group, with context
Answers where employees already work — the cheapest rollout.
06
Continuous improvement
Know what went unanswered
Every turn is judged answered or not; gaps are listed per base with what is missing and how it was searched. Dashboards, run logs, usage and audit show quality and cost.
Knowledge gapsUnanswered questions with asker, missing info and documents found
Q&A dashboardVolume, completion, estimated cost and trends by tool and base
Run logsEvery step and every retrieval can be expanded
Usage & auditConsumption per person and model; who did what and when
The knowledge base gets more complete the more it is used.
Product tour
Screenshots from a demo environment with a fictional company
Specifications
Self-hosted, using your existing models and infrastructure
Deployment
One-command Docker ComposeLinux x86_644 cores / 8 GB to startOffline installAutomatic DB upgrades
Models
OpenAI-compatible APIsDeepSeek / Qwen / self-hosted vLLMChat, embedding, rerank, visionThinking switch
Documents
PDFWord (.docx / .doc)PPTExcel / CSVMarkdown / text200 MB per file
Sources
UploadS3-compatible storageSMB sharesWebDAVScheduled syncExport / import
Enhanced retrieval (optional)
Multi-hop & knowledge graphVisual retrieval (ColPali)Structure enhancement
Login & APIs
Password & invitesOIDC SSOPersonal API keysRESTMCPSwagger docs
Compared with a generic AI chat tool
The difference is everything around the model
| KnowWork | Generic AI chat tool | |
|---|---|---|
| Where data lives | On your servers; documents, chats and vectors stay inside | Questions and files go to the vendor |
| Company knowledge | Built-in base with permissions and version detection | Upload files each time, or connect a base without permissions |
| Sources | Citations open the original paragraph | Usually not verifiable |
| Access control | Spaces, departments, groups, folder grants | Whoever can use it can ask everything |
| Taking action | Sandboxed, writes need approval | Limited to the vendor’s tools |
| Visibility | Gaps, dashboard, run logs, audit | Little insight into what was asked and how it went |
Open source and commercial
Open source ships the full agent and permissions; commercial adds the built-in knowledge base and workspaces
| Capability | Open source | Commercial |
|---|---|---|
| Chat, session sandbox, Office files | ✓ | ✓ |
| Skills, MCP tools, approval before writes | ✓ | ✓ |
| Members, departments, groups, grants, OIDC | ✓ | ✓ |
| WeCom, Feishu and DingTalk bots | ✓ | ✓ |
| Dashboard, run logs, usage and audit | ✓ | ✓ |
| Search API and MCP | ✓ | ✓ |
| Connect KnowFlow knowledge base | ✓ | ✓ |
| Built-in knowledge base with versions and gaps | — | ✓ |
| Multiple workspaces and sharing | — | ✓ |
| Multi-hop, visual retrieval, structure enhancement | — | ✓ |
Typical scenarios
Start with one department and one kind of question
Policy Q&A
Leave, expenses, procurement — answered any time with the policy text, newest version first.
HRFinanceQuick answer
Contract & bid review
Upload a contract or bid, check it against company rules, get the risks as a table.
SkillsAttachmentsWord / Excel
Sales & support assistant
Manuals, pricing and cases in one place, asked directly from chat apps.
Chat botsProduct docsCitations
Group-wide deployment
One space per subsidiary, each managing its own; group policies shared to all.
WorkspacesSharingFolder grants
Try it with your own documents
We provide the package and POC support to run it on your servers, with your models and files