MultiMind SDK: Features & Functions

For the authoritative, per-feature status (Stable / Beta / Planned), see feature-status.md. This page gives a structural overview; items marked “Planned” below are not implemented yet.

​

Project Structure

multimind-sdk/
├── multimind/
│   ├── __init__.py                # Public API; heavy features are lazy-imported per extra
│   ├── _lazy.py                   # Lazy attribute loader used by the package
│   ├── main_config.py             # Central config loader
│   ├── config/                    # Config helpers
│   ├── core/                      # Gateway config, chat sessions, monitoring, exceptions
│   ├── models/                    # Model wrappers: openai.py, claude.py, ollama.py, huggingface.py,
│   │                              #   groq.py, gemini.py, mistral.py, deepseek.py, openrouter.py,
│   │                              #   together.py, xai.py, perplexity.py, fireworks.py, cerebras.py,
│   │                              #   multi_model.py, factory.py, moe/
│   ├── llm/                       # Non-transformer LLMs (Mamba, RWKV, scaffolds), model registry
│   ├── router/                    # router.py, strategy.py, fallback.py, adaptive.py, multi_modal_router.py
│   ├── rag/                       # rag.py (RAG, RAGConfig), base.py, fluent.py, graph_retrieval.py, postprocessing.py
│   ├── retrieval/                 # Retriever, EnhancedRetriever, HybridRetriever
│   ├── embeddings/                # EmbeddingConfig, EmbeddingModel
│   ├── vector_store/              # VectorStore + ~60 backends (FAISS, Chroma, Qdrant, ...)
│   ├── document_loader/           # File/data ingestion
│   ├── document_processing/       # Chunkers and processors
│   ├── fine_tuning/               # lora_trainer.py, qlora_trainer.py, PEFT and meta-learning modules
│   ├── agents/                    # agent.py, memory.py, agent_loader.py, agent_registry.py, tools/ (calculator.py)
│   ├── memory/                    # ~50 memory classes (see memory.md)
│   ├── orchestration/             # prompt_chain.py, task_runner.py
│   ├── mcp/                       # Model Composition Protocol: parser.py, executor.py, schema.json
│   ├── mcp_server/                # Anthropic Model Context Protocol server
│   ├── ensemble/                  # AdvancedEnsemble and ensemble methods
│   ├── compliance/                # PII guard, privacy, audit, regulation checkers
│   ├── integrations/              # github.py, slack.py, discord.py, jira.py, model_adapters.py, frameworks/
│   ├── multimind_logging/         # trace_logger.py, usage_tracker.py
│   ├── observability/, metrics/   # Tracing and cost metrics
│   ├── gateway/                   # FastAPI apps: api.py, rag_api.py, compliance_api.py, guard_proxy.py
│   ├── api/                       # multi_model_api.py, unified_api.py
│   ├── server/                    # MultiMindServer base app and platform.py (combined "MultiMind API")
│   ├── dashboard/                 # Governance dashboard
│   ├── client/                    # RAGClient, ModelClient family, FederatedRouter, ModelSession
│   └── cli/                       # click CLI: chat, models, config, compliance, audit, serve, dashboard, backend
├── examples/                      # Example scripts (see examples/README.md)
├── tests/                         # Test suite
├── docs/                          # Documentation
├── README.md
├── LICENSE
├── pyproject.toml
└── setup.py

​

Core Features

​

Model Support

• Model Wrappers (available today):
  • OpenAI, Anthropic Claude, Ollama (local models, including Mistral and Llama), HuggingFace (Inference API or local transformers)
  • Groq, Google Gemini, Mistral AI, DeepSeek, OpenRouter, Together, xAI, Perplexity, Fireworks, Cerebras
  • MultiModelWrapper for primary + fallback models
• Unified Interface: All wrappers derive from BaseLLM and support generate, chat and streaming (generate_stream, chat_stream); embeddings where the provider offers it
• Non-Transformer Inference: Mamba and RWKV run via HuggingFace-backed wrappers; other listed architectures are extension scaffolds that raise NotImplementedError

​

Agent System

• Agent Framework (available today):
  • Agent(model, memory, tools, system_prompt, max_turns) with native function calling on OpenAI-compatible models and keyword-based tool routing elsewhere
  • AgentMemory for task/response history
  • Agent configuration loading from MCP workflow files (agent_loader)
  • One built-in tool (CalculatorTool); extensible via BaseTool subclasses
  • AgentOrchestrator, SelfEvolvingAgent, ReasoningChain (ReAct-style steps)
• Planned: hierarchical multi-agent systems, inter-agent communication, AutoML-based agent configuration, additional built-in tools (web search, file operations)

​

Orchestration

• Prompt Chains: multi-step workflows with variable substitution and conditional execution
• Task Runner: task dependencies, retry logic, context management, parallel execution

​

Model Composition Protocol (MCP)

Note: multimind.mcp is MultiMind’s own Model Composition Protocol, a JSON/YAML workflow executor for chaining models. Anthropic’s Model Context Protocol is supported separately by multimind.mcp_server (pip install "multimind-sdk[mcp]", python -m multimind.mcp_server; see mcp-server.md).

• Workflow Definition: JSON/YAML workflow specification validated against schema.json; model composition and chaining; conditional execution; data transformation
• Execution Engine: workflow validation, model registration, step execution, error handling (executor.py, advanced_executor.py)

​

Routing & Strategy

• Model Router: strategy-based routing (cost-aware, latency-aware, hybrid), fallback chain, adaptive and multi-modal routers
• Federated routing: multimind.client.FederatedRouter chooses local vs. cloud models by input size, latency and privacy

​

Fine-Tuning & PEFT

• Available today (basic): LoRA and adapter training; QLoRA (real 4-bit NF4 + LoRA when the [finetune-gpu] deps are installed); synthetic Q/A generation for RAG fine-tuning (SyntheticQAGenerator, no torch needed)
• Experimental / partially implemented: IA³, BitFit, MAM adapters, UniPELT, meta-learning (MAML, Reptile), adaptive PEFT
• Planned: HyperLoRA, Compacter, AutoML-based parameter tuning, advanced optimization

​

RAG (Retrieval-Augmented Generation)

• Vector stores: FAISS and Chroma are the most exercised; Pinecone, Qdrant, Weaviate, Milvus, Zilliz, Elasticsearch, OpenSearch, MongoDB Atlas, the PostgreSQL family and ~25 other client-backed stores are implemented; eleven backends are NotImplementedError stubs (full list in feature-status.md)
• Retrieval: RAG.retrieve, EnhancedRetriever, HybridRetriever, GraphRetriever (entity-graph expansion), post-processing, fluent RAGPipeline
• Known limitation: RAG.retrieve(filter_criteria=...) currently raises; filter on result metadata instead (see rag.md)
• Planned: self-tuning indexer, quantum-enhanced search

​

Enterprise Integration

• Compliance Suite (available today): PII detection and redaction, GDPR/HIPAA policy modeling, hash-chained audit logging, cost chargeback, OpenAI-compatible guard proxy (multimind serve), compliance dashboards (with the [compliance] extra)
• Service Integrations (available today): GitHub, Slack, Discord, Jira adapters in multimind.integrations; framework adapters for LangChain, LlamaIndex, CrewAI, AutoGen and Haystack in multimind.integrations.frameworks (each has its own extra)
• Planned: edge deployment toolkit, plugin system (Notion, Salesforce), database connectors, real-time data integration (Kafka, MQTT)

​

Development Tools

• Available today: governance dashboard (multimind dashboard), shadow-AI inventory scan (multimind audit scan), combined Swagger UI (multimind backend)
• Planned: visual workflow builder, drag-and-drop components, no-code interface

​

Logging & Monitoring

• Usage Tracking: token usage, cost tracking, operation logging (multimind_logging.usage_tracker)
• Trace Logging: request/response tracing, performance metrics, error tracking (multimind_logging.trace_logger)
• Gateway metrics: per-model request counts, latency, error counts, health checks and rate limits (multimind.core.monitoring), exposed at GET /v1/metrics
• Export & Reporting: JSON/CSV export, usage summaries, cost chargeback (multimind audit costs)

​

Configuration

• YAML/JSON config files, .env and environment variable support (python-dotenv is loaded automatically)
• CLI defaults in ~/.multimind_cli_config (multimind config manage)

​

CLI & Integration

• CLI (multimind): chat, models (list/compare/health/metrics/download/export/delete), config, compliance, audit, serve, dashboard, backend, convert — see cli.md
• Service Integrations: GitHub, Slack, Discord, Jira; model adapter classes for custom model types

​

Feature Table

​

Examples

• Basic Agent Usage - Agent creation and usage
• Prompt Chaining - Multi-step reasoning workflows
• Task Running - Workflow orchestration
• MCP Workflows - Model composition examples
• Usage Tracking - Monitoring and logging

​

Architecture Overview