Features

OpsPilot organizes the AI building and management pipeline into three layers: underlying resources, mid-level logic processing, and top-level view delivery. Below is a breakdown of key features by platform navigation module.

1. Models (Provider)

As the foundation of a full-stack agent, "Model Management" is designed to centrally register, store, and manage LLM API assets from various vendors within the enterprise.

• Multi-Dimensional Model Governance: Supports centralized integration and toggling of various model categories including LLM, Embedding, Rerank, and even OCR (eliminating the need for business users to repeatedly configure LLM environments and authentication for each project).
• Controllable Visibility Boundaries: Each model supports isolation management through "team groups" and status controls (addressing the risks of disordered model assets and cross-department authorization breaches).
• Sensitive Configuration Encryption: Backed by high-security logic controls, fields such as API Keys are automatically encrypted (preventing plaintext password exposure in the frontend or log leakage).
• Vendor Protocol Types: Vendors support selecting between "OpenAI Compatible" and "Anthropic Compatible" protocol types, meeting the requirement for mixed deployment of different protocol models under the same API gateway.

2. Tools (Tool)

Tool management extends LLMs beyond passive question-answering, enabling them to "reach out" and interact with your existing internal enterprise systems.

• Out-of-the-Box Integration Architecture: Supports one-click creation from a built-in library or custom import of invocation schemes (enabling LLMs to directly handle tasks such as network monitoring, ticket approval queries, and more).
• MCP Protocol Extension: Strong compatibility with MCP tool fetching capabilities at the underlying layer (facilitating users to integrate a vast array of third-party ecosystem APIs through the latest protocol).
• Context Parameter Isolation: Tool execution parameters feature clear "text" or "password" differentiation (ensuring parameters are valid and sensitive data flows securely when tools send instructions to external third parties).

Skill Packages (SkillPackage) (Beta)

Skill Packages are independently packaged and distributable agent method collections, encapsulating domain-specific knowledge (SKILL.md), execution strategies (skill.yaml), and dependent tool declarations in a standardized ZIP format, which can be imported into any agent to activate domain capabilities on-demand.

• ZIP Import: Upload ZIP files conforming to specifications in "Tools → Skill Packages"; the platform automatically parses manifests, validates security policies (preventing custom code execution), and performs idempotent updates using a three-tuple key of package_id + version + domain.
• Trigger Word Auto-Matching: The runtime engine reads the triggers field from the skill package; when user messages match trigger words, the corresponding skill package's SKILL.md guide is automatically injected into the agent's system prompt, guiding the LLM to execute tasks according to established procedures.
• Dependent Tool Declaration: required_tools lists the names of necessary tools; the runtime checks tool availability in real-time and annotates missing items in the prompt, preventing the LLM from blindly attempting execution when tools are unavailable.
• Team Group Visibility Control: Each skill package belongs to one or more team groups; members can only view skill packages in their assigned groups, isolating capability leakage across business domains.
• Enable/Disable Management: Skill packages can be individually enabled or disabled; when disabled, they are no longer injected into prompts but do not affect the normal operation of mounted tools.

3. Knowledge Base (Knowledge)

The knowledge base is the definitive remedy for mitigating the "hallucination" risk of large language models. Enterprises can centrally manage their "memory" here.

• Full-Format Text Content Extraction: Supports knowledge sources via direct local document upload or custom links (addressing the inability of LLMs to answer domain-specific questions due to lack of real-time access to specialized corpora). Supported file types include: md, docx, xlsx, csv, pptx, pdf, txt, and png, jpg, jpeg (with OCR model required).
• Automatic Preprocessing & Visual Cleansing: The backend handles document chunking and segmentation steps (resolving the severe limitations faced by non-AI-specialist administrators when adjusting retrieval chunk sizes).
• Processing Status & Error Display: After documents enter the processing queue, the platform displays real-time parsing progress and training percentage; if processing fails, error messages are displayed directly in the document list, enabling operations teams to identify issues without checking backend logs.
• Fine-Grained Recall Control: Supports dynamic changes to retrieval Rerank models, with configurable recall modes and similarity threshold filtering (helping system administrators optimize the accuracy and confidence of AI-adopted answers).
• Relational Graph Restructuring: Supports knowledge graph extraction and generation based on Q&A documents (enabling LLMs to perform clearer causal and panoramic structural logical profiling).

4. Agents (Skill)

Agents are logical unit carriers that assign "execution roles" to LLMs in specific scenarios.

• Ready-to-Use Scenario Templates: Built-in multi-category agent template library (addressing the pain point of beginners not knowing how to tune Prompt parameter settings when facing complex parameter configurations).
• Plug-and-Play Enhancement Mounting: Directly enable RAG capabilities and bind associated "Knowledge Bases" with adjustable threshold parameters (ensuring agents can think within defined scope using tightly coupled business logic).
• Long-Term Memory Management: Provides strict mode and chat history carry-over settings (enabling agents to maintain contextual conversation continuity tracking over extended interaction timelines). Memory capabilities are managed through an independent "Memory Space (MemorySpace)" module, with a local storage engine currently deployed; Mem0, Zep, and custom API are reserved storage-type enum options without usable engine implementations yet—refer to future releases for updates.

5. Studio & ChatFlow

The mid-platform workspace where delivery personnel ultimately assemble agents and deliver services.

• Service Mode Switching: Provides application cards for single-scenario Pilot customization and LobeChat session-type architecture management (addressing engineering obstacles preventing modular deployment of intelligent applications).
• Multi-Source Trigger Access: Built-in rich workflow entry capabilities, aligned with OpenAI parallel interfaces, RESTful APIs, and scheduled batch processing (expanding upstream integration channels).
• WYSIWYG Graph Orchestration: Controlled by a topology canvas built on node-based ChatFlow.
• Test Execution vs. Real Conversations: Debug executions initiated from the Studio canvas are marked as "test execution" (is_test=true), strictly isolated from normal conversations initiated through channels, with only administrative organization members allowed to trigger test executions. Execution records are separately marked to prevent debug data from contaminating production logs and analytics.
• Workflow Attachment Assets: The workflow's "Attachment Generation" node (attachment_file tool) can generate downloadable files in a single execution, currently supporting Markdown, PDF, and Word (docx) formats. Each attachment is associated with a unique signed download token, enabling channels (email/enterprise messaging) to deliver them directly to users as attachments.
• Bot Dual-Layer Organization Authorization (Beta): Bots adopt a "Managing Organization / Usage Organization" dual-layer authorization model. The managing organization has complete administrative permissions (view, edit, delete, authorize) and inherently possesses conversation usage rights. Administrators can additionally grant conversation permissions to one or more "Usage Organizations" in the Studio, enabling members of those organizations to initiate conversations through various access channels, while they do not possess administrative capabilities. The managing organization is always included in the scope of usage organizations and cannot be removed. Conversation authorization is scenario-dependent: Test Execution within the Studio is restricted to managing organization members only; Normal Conversations initiated through access channels are open to all members of authorized usage organizations (enabling cross-organization service sharing while ensuring the managing team maintains exclusive and indivisible control over the bot).
• Human-AI Collaboration Interface Authorization Hardening (Beta): Approval and selection operation interfaces in human-AI collaboration workflows have been upgraded from permissionless access to mandatory credential verification, with concurrent validation of the operator's organization affiliation during the verification phase—only members belonging to the bot's configured organization scope can proceed through operation requests, preventing external or unauthorized calls from interfering with approval workflows.

Interface Guide:

• Configuration Logic: On this canvas, if you need to design logic "branches," simply drag a "Logic Decision" node from the left panel, and route the flow to different agents and actions based on the Q&A intent context.

6. Session Tracking & Platform Channels (Sessions)

Operations and compliance support components.

• Multi-Channel Multimodal Publishing: Supports open-loop integration with DingTalk, WeChat Official Accounts, WeCom groups, and even GitLab (addressing the difficulty of integrating model platforms into enterprise collaboration environments).
• Business Audit Closed-Loop Logging: Provides visual interactive Chat Logs and full-chain usage details for underlying compute/token consumption (helping enterprises mitigate business risks and control invisible model consumption costs).

Warning / Security Best Practices: To protect data assets throughout the entire process, be sure to set appropriate team scopes for "Knowledge Bases" and associated "Agents" to limit cross-contamination. Additionally, all external platform callback Webhook URLs should be stored with strong desensitization.