casibase

Abstracts 30+ AI model providers (OpenAI, Claude, Gemini, Llama, Ollama, HuggingFace) behind a single chat API using a pluggable provider registry pattern. Routes chat requests to configured providers via standardized adapter interfaces, handling model-specific parameter mapping, streaming responses, and error fallback. Implemented via provider.go model with provider-specific controller logic that normalizes request/response formats across heterogeneous APIs.

Implements a retrieval-augmented generation pipeline that embeds documents into vector space using configurable embedding providers, stores vectors in a knowledge base (Store entity), and retrieves semantically similar documents during chat to augment LLM context. The system uses vector.go to manage embeddings, store.go for knowledge base configuration, and integrates with the AI answer generation pipeline to inject retrieved context into prompts before sending to LLMs.

Provides email notifications for chat events (new messages, mentions), workflow completions, and system alerts. Integrated with the message lifecycle (message.go) and background task system (main.go), allowing notifications to be triggered based on configurable rules. Email provider is abstracted through the provider system, supporting multiple SMTP backends and email service providers.

Implements specialized features for medical applications including electronic health record (EHR) integration, HIPAA-compliant data handling, and medical document parsing. Medical records are stored with enhanced encryption, access control is audit-logged, and sensitive data is masked in logs. Integrated with the knowledge base system for medical document indexing and the security scanning system for compliance validation.

Provides integration with Kubernetes for deploying Casibase and managing containerized AI workloads. Includes Helm charts, deployment manifests, and orchestration logic for scaling chat services, managing provider connections, and handling stateful components (databases, vector stores). Deployment configuration is managed through the application configuration system (conf/app.conf) with environment-based overrides for different Kubernetes clusters.

Implements comprehensive internationalization using a JSON-based locale system (web/src/locales/en/data.json, web/src/locales/zh/data.json) supporting multiple languages. All UI strings are externalized to locale files, allowing language switching without code changes. Backend supports locale-aware responses (timestamps, number formatting) and the frontend dynamically loads locale data based on user preference.

Implements graph visualization capabilities (graph visualization system in web/src/App.js) for exploring relationships between documents, entities, and concepts in the knowledge base. Supports interactive graph rendering, node/edge filtering, and traversal. Integrated with the knowledge base system to automatically extract and visualize entity relationships from indexed documents.

Provides content management for articles and workflows, with built-in analytics tracking user interactions, chat usage, and knowledge base access patterns. Analytics data is collected via event tracking in the frontend and backend, aggregated in the database, and visualized in dashboards. Supports custom metrics and event definitions for domain-specific analytics.

Implements Model Context Protocol support enabling Casibase to act as an MCP server and client, allowing external tools and agents to be registered as context providers. The system uses a schema-based function registry (via provider system) to expose capabilities to connected clients, handle tool invocation requests, and manage bidirectional agent-to-agent (A2A) communication. MCP integration is managed through the provider ecosystem, allowing tools to be dynamically registered and invoked during chat.

Provides built-in user management integrated with Casdoor (external identity provider) for Single-Sign-On, role-based access control (RBAC), and permission management across knowledge bases, chats, and admin functions. Implemented via authz_filter.go and auto_signin_filter.go middleware that intercepts requests, validates user identity and permissions, and enforces access policies. Supports multi-tenant scenarios with per-user knowledge base visibility and chat history isolation.

Supports uploading diverse document types (PDF, markdown, text, web content) into a knowledge base (Store entity), automatically parsing and chunking documents, generating vector embeddings via configured embedding providers, and storing vectors for semantic search. The system uses file_cache.go and local_file_system.go for storage abstraction, integrates with document parsers (initialized in main.go), and manages the Store entity lifecycle including file tree permissions and access control.

Implements streaming responses from LLMs using provider-agnostic adapters that normalize streaming APIs across different providers (OpenAI, Claude, Ollama, etc.). The chat pipeline uses server-sent events (SSE) or WebSocket to stream tokens to the frontend in real-time, with provider adapters handling format conversion (e.g., OpenAI's delta format to generic token stream). Message lifecycle is managed via message.go model with transactional consistency and retry logic for failed streams.

Provides a React-based admin dashboard (ProviderEditPage.js, ProviderListPage.js, StoreBackend.js) for managing provider credentials, knowledge bases, embeddings, and system configuration without code changes. The UI communicates with backend controllers (provider.go, store.go, vector.go) via REST APIs, allowing admins to add/remove providers, configure embedding settings, upload documents, and manage access control. Configuration is persisted in the database and hot-reloaded without server restart.

Implements a specialized workflow for video annotation, review, and asset management (VideoBackend.js, video annotation system in web/src/App.js). Supports uploading videos, frame extraction, annotation markup, and collaborative review. Integrated with the asset scanning system for security validation and metadata extraction. Videos are stored via the provider storage abstraction and indexed for search.

Implements automated security scanning of uploaded assets (documents, videos, images) using configurable scanning providers. The system validates file types, scans for malware/sensitive content, and enforces compliance policies before indexing. Scanning is integrated into the ingestion pipeline (asset scanning system in web/src/App.js) and can block uploads that fail validation. Results are logged for audit purposes.

Implements remote access tunneling (tunnel.go) enabling secure, encrypted communication between Casibase instances and external agents/services without exposing ports. Uses a tunnel protocol to establish bidirectional communication channels, allowing agents to access Casibase APIs and knowledge bases securely. Tunneling is managed via the provider system and supports authentication/authorization for tunnel endpoints.