RatioMax Enterprise AI Knowledge Base & Agent Platform: Bring your documents to life

Positioning

From static document storage to living knowledge

Traditional document systems only store and search. RatioMax is the intelligence layer on top of enterprise knowledge management — turning static files into dynamic assets that AI can call and business workflows can integrate.

• Storage: Traditional document systems — Upload, folders, versions; RatioMax · Living knowledge — Multi-source sync, multi-format parsing, structured ingestion
• Retrieval: Traditional document systems — Keyword full-text search; RatioMax · Living knowledge — Semantic vectors + keywords hybrid search + reranking
• Structuring: Traditional document systems — None; RatioMax · Living knowledge — Agents autonomously build Wiki pages and knowledge graphs
• Application: Traditional document systems — Manual lookup; RatioMax · Living knowledge — Multi-step agent reasoning, tool calls, multi-channel service
• Service: Traditional document systems — No standard interface; RatioMax · Living knowledge — MCP protocol · 29 tools · 3 transports

Core value

Three dimensions that reshape enterprise knowledge

Knowledge as assets

• Scheduled sync from Feishu Wiki / Drive, Notion, Yuque and RSS; Chrome extension capture; API / CLI upload
• Parses 10+ formats including PDF, Word, images, Excel, EPUB and MHTML
• Agents autonomously generate Wiki pages and knowledge graphs with cross-document links
• Visual chunk editing, version snapshots, diffs, one-click rollback and auto re-indexing

Intelligent services

• ReAct (reason + act) loops drive multi-step tasks with parallel tool calls and thinking modes
• Hybrid semantic + keyword retrieval with reranking and cross-base search for higher precision
• A dedicated data-analysis agent for querying and analysing data
• Langfuse tracing covers the full path from question to retrieval

Standardized integration

• MCP interface: 29 tools over stdio / SSE / HTTP with OAuth2 and human approval
• WeCom, Feishu, DingTalk, Yunzhijia, Slack, Telegram and more IM channels
• 20+ model vendors including DeepSeek, Qwen, Zhipu, Hunyuan, OpenAI and Gemini
• Direct file links let third-party apps load files without re-authentication

Data flow

From a single document to a cited answer

Data moves through four stages — each one observable and debuggable.

1. Ingest: Documents arrive via API, IM or connectors; an ingestion job is created and queued.
2. Structure: Parsers extract text, chunk it by strategy, then embed it into vector and object storage.
3. Retrieve: Questions are embedded for semantic search, merged with keyword search and reranked.
4. Apply: Results enter the model context; the agent reasons, calls tools and answers with citations.

Capabilities

One platform for knowledge and agents

From parsing, chunking and retrieval to reasoning and structuring — every layer is designed for enterprise work.

RAG engine: Retrieve enterprise knowledge first, then answer

Chunking decides RAG quality. RatioMax provides multi-level chunking with full lifecycle management, hybrid retrieval with reranking, and automatic citations — click to see the original paragraph.

• Multiple parsers: OpenDataLoader for complex PDFs, PaddleOCR-VL for images and scans
• Live debugging panel to preview chunking parameters instantly
• Version snapshots, diffs, one-click rollback and custom metadata filters
• Fan-out across 7 vector DBs: pgvector, Milvus, Qdrant, Weaviate, OpenSearch, Apache Doris, Tencent Cloud VectorDB

Agents & tools: Enterprise agents that reason and act

Agents run ReAct loops to plan steps, call tools, observe results and adapt. Independent tools run in parallel within a single round to cut reasoning turns.

• Parallel tool calls complete complex tasks in fewer rounds
• Skill system with preset skills
• Sandboxed, scoped execution prevents privilege escalation
• Dedicated data-analysis agent for querying data

Wiki structuring: Turn scattered files into a linked knowledge network

Wiki mode is a RatioMax differentiator. Agents generate structured, interlinked Markdown Wiki pages and knowledge graphs from raw documents — far beyond simple summaries.

• Extract → generate pages → cross-reference → knowledge graph
• Page snapshots, line-level diffs, rollback and in-browser editing
• Task queue with dead-letter queue handles batches of tens of thousands of documents
• Entities and relations visualised at a glance

Observability: Every answer can be traced, replayed and improved

Langfuse and W3C cross-service tracing record every step from question to retrieval, while task scheduling is built for high concurrency and slow jobs.

• Staged worker pools keep parsing, embedding and reranking independent
• Per-model concurrency limits protect model API rate limits
• Failed jobs land in a dead-letter queue to inspect, retry or discard
• Prompt cache hit rate and token savings at a glance

Customers

In production across advanced materials and equipment manufacturing

Results shown are project targets

Security & deployment

The stricter your compliance, the more you need private deployment

Deploy the full platform on your own infrastructure with complete data control. Permissions, keys, identity and storage are all built for defense in depth.

• Space-level permissions: Space isolation with a four-tier role matrix (RBAC), down to individual knowledge bases, with per-space audit logs.
• Two-tier API keys: Fine-grained keys (capability + knowledge-base scoped) and platform keys for control-plane automation.
• Single sign-on: Connect your enterprise SSO via OIDC and keep your existing identity system.
• Defense in depth: Hardened transport, storage, API and app layers; MCP and connector credentials encrypted with AES-256-GCM; SSRF protection built in.

Flexible delivery, elastic scale

• Docker Compose: Single-node deployment for fast validation
• Kubernetes: Production-grade cluster deployment
• Microservices: Four core services deploy and scale independently over encrypted channels and queues
• Multi-instance storage: Bind storage per knowledge base: classified data on private OSS, public data on S3

Ecosystem

Fits into the tools and infrastructure you already run

Mix Chinese and international models freely, connect mainstream IM channels, and use the storage and vector databases of major cloud providers — compose what you need.

IM channels

WeCom, Feishu, DingTalk, Yunzhijia, WeChat, QQBot, Slack, Telegram, Mattermost

Model vendors

DeepSeek, Qwen, Zhipu, Hunyuan, MiniMax, OpenAI, Gemini, Anthropic, Azure OpenAI, NVIDIA, Ollama

Vector DBs

pgvector, Milvus, Qdrant, Weaviate, OpenSearch, Apache Doris, Tencent VectorDB

Data sources

Feishu Wiki, Feishu Drive, Notion, Yuque, RSS, WeChat Mini Program, Chrome extension

MCP standard tool protocol

Published as the official PyPI package ratiomax-mcp, covering knowledge-base management, retrieval, Wiki operations and sessions. Third-party apps obtain tokens via standard OAuth2, and sensitive actions require human approval.

• Chrome extension: Capture selected text, images or whole pages as knowledge entries
• WeChat Mini Program: Ask your knowledge base directly inside WeChat
• WeChat dialog platform: Zero-code Q&A for Official Accounts and Mini Programs
• Embeddable widget: Secure token exchange and rate limiting to publish agents on external sites

Where we fit

When we may not be your best fit

• You mainly need a searchable file repository: If the goal is storing, organising and keyword-searching documents, a general-purpose cloud drive or your OA attachment library is usually lighter and cheaper, and easier to fit into existing workflows.
• You only need a general-purpose chat assistant: If day-to-day usage is general Q&A without enterprise-specific knowledge or citations, a public-cloud general-purpose LLM product is cheaper to start with.
• You are not ready for on-premise deployment and operations: RatioMax is built for organisations willing to self-host and keep data under their own control. If you would rather not take on operations, an out-of-the-box managed product will suit you better.
• You want conclusions only, without traceable citations: If answers can be taken at face value without showing their sources, our citations and traceability may not be essential — they simply add steps.
• You have no plans to embed knowledge into your own systems: If knowledge is used only inside the platform and never through standard interfaces into your existing systems, our MCP / API integration offers limited value.
• Budget or timeline cannot yet support knowledge governance: Knowledge governance needs ongoing curation effort. If that is hard to arrange now, start with a small pilot in one team or one document type, then scale once it proves out.

Capability timeline

Continuous, frequent iteration

• Agent mode launched
• Shared spaces & skills
• All channels: WeCom, Feishu, Slack, TG, DingTalk
• Cloud & ecosystem
• Wiki knowledge structuring
• 10k-document scale & debugging
• Space-level RBAC
• Dual parsers: OpenDataLoader + PaddleOCR-VL
• Permissions & operations
• Audit & tracing: Langfuse end-to-end
• Docs site & chunk governance

About us

Making every enterprise document something you can talk to, reason with and reuse

RatioMax (汇理极知) builds an enterprise AI knowledge base and agent platform. It is in production in modified engineering plastics and mining machinery, covering knowledge assets, intelligent Q&A, Wiki structuring and multi-channel integration.

Data-driven · Smart future

Industries

What a knowledge base actually solves, industry by industry

A platform page explains capabilities. An industry page explains problems. Each page below breaks down what knowledge looks like in that sector, where it gets stuck, and which parts of RatioMax connect to it.

• Wear Parts & Mining Equipment: AI Knowledge Base for Wear Parts
• Equipment Manufacturing & Drawings: Drawing versions, process and inspection knowledge
• Modified Engineering Plastics: Engineering Plastics R&D and Quality Knowledge Base
• Government & Public Sector: AI Knowledge Base for Public Services
• Financial Services: AI Knowledge Base for Financial Services Teams

FAQ

Questions enterprises ask about RatioMax

What is RatioMax?

RatioMax (汇理极知) is an enterprise AI knowledge base and agent platform. It parses, chunks and embeds unstructured content scattered across drives, documents and IM, then combines retrieval-augmented generation (RAG), ReAct agent reasoning and auto-generated Wiki knowledge graphs to deliver grounded answers with traceable sources.

Does RatioMax support private (on-premise) deployment?

Yes. The full platform can run on your own servers or intranet with complete data control. Deploy with Docker Compose on a single node or Kubernetes for production clusters; four core services deploy and scale independently, and each knowledge base can bind to private OSS or S3 storage.

How does RatioMax keep AI answers accurate and traceable?

It retrieves enterprise knowledge before answering, using hybrid semantic and keyword search with reranking. Every answer carries citations that open the original paragraph, and Langfuse plus W3C cross-service tracing record each step from question to retrieval.

Which LLMs are supported? Can we use Chinese models?

RatioMax supports 20+ model vendors, including Chinese models such as DeepSeek, Qwen, Zhipu, Tencent Hunyuan and MiniMax, as well as OpenAI, Gemini, Anthropic, Azure OpenAI and NVIDIA. Local models can be connected through Ollama.

Can it connect to WeCom, Feishu and DingTalk?

Yes. Channels include WeCom, Feishu, DingTalk, Yunzhijia, WeChat, QQBot, Slack, Telegram and Mattermost, plus a WeChat Mini Program, Chrome extension and embeddable website widget.

Which document formats and data sources are supported?

RatioMax parses 10+ formats including PDF, Word, Excel, images, EPUB and MHTML — OpenDataLoader handles complex PDFs and PaddleOCR-VL handles images and scans. It syncs on schedule from Feishu Wiki / Drive, Notion, Yuque and RSS, and accepts Chrome extension capture and API / CLI uploads.

How does RatioMax handle knowledge governance?

Visual chunk editing, a live debugging panel, version snapshots, diffs and one-click rollback with automatic re-indexing, plus custom metadata filters. Wiki mode lets agents generate interlinked pages and knowledge graphs from raw documents so knowledge stays structured and maintainable.

How are data security and permissions handled?

Space isolation with a four-tier RBAC role matrix down to individual knowledge bases, per-space audit logs, OIDC single sign-on, two-tier API keys, AES-256-GCM encryption for MCP and connector credentials, and built-in SSRF protection.

What is MCP and how does RatioMax expose knowledge services?

MCP (Model Context Protocol) is a standard protocol connecting AI applications to external tools. RatioMax publishes the official PyPI package ratiomax-mcp with 29 standard tools over stdio / SSE / HTTP, OAuth2 authorization and human approval for sensitive actions, alongside roughly 360 documented API endpoints.

Which enterprises and industries is RatioMax for?

Organisations with strict data compliance that need private deployment — such as finance, government and healthcare — and teams that need multi-step agents, structured knowledge and multi-channel service. RatioMax is in production in modified engineering plastics (Guangdong Zhongsu New Materials, 301686) and mining machinery (Chongqing Boyue Machinery).

Contact

Book a tailored demo

Tell us what you need. Our solutions team will get in touch to assess private deployment and a practical rollout path for your business.

Sales: system@agent99.cn

Support: system@agent99.cn