Minima

Automatically discovers and processes documents across multiple formats (.pdf, .xls, .docx, .txt, .md, .csv) from a configured local directory tree, extracting text content and preparing it for embedding generation. Uses recursive folder traversal to handle nested directory structures without manual file selection, enabling hands-off indexing of large document collections.

Generates dense vector embeddings for document chunks using Sentence Transformers (BAAI models by default), converting text into high-dimensional vectors suitable for semantic similarity search. Supports model selection via environment configuration, allowing users to choose embeddings optimized for their domain (e.g., multilingual, domain-specific fine-tuned models) without code changes.

Stores generated embeddings in Qdrant vector database and performs approximate nearest neighbor (ANN) search to retrieve semantically similar documents for a given query. Uses vector similarity metrics (cosine, Euclidean) to rank documents by relevance without keyword matching, enabling natural language search across document collections.

Applies a second-stage ranking model (typically BAAI cross-encoder) to refine the top-k results from vector search, re-scoring documents based on semantic relevance to the original query. This two-stage retrieval pattern (retrieve-then-rerank) improves precision by filtering out false positives from the initial ANN search without requiring full dataset re-scoring.

Abstracts LLM interaction behind a provider interface supporting Ollama (local), OpenAI (ChatGPT), and Anthropic (Claude) without code changes. Uses environment configuration to select the active LLM backend, enabling users to switch between fully local inference and cloud LLMs based on deployment mode, privacy requirements, or cost considerations.

Exposes Minima's RAG capabilities as a Model Context Protocol (MCP) server, allowing external LLM clients (Claude Desktop, other MCP-compatible applications) to invoke document search and retrieval as remote tools. Implements MCP's request-response protocol for tool discovery, invocation, and result streaming without requiring direct API integration.

Provides dual user interfaces for document search and RAG interaction: a web-based UI (accessible via browser) and a native Electron desktop application. Both interfaces connect to the same backend services (indexer, vector database, LLM) and support chat-style interaction with retrieved context, enabling non-technical users to search documents without CLI or API knowledge.

Centralizes all system configuration through environment variables (.env file), including document paths, embedding models, vector database endpoints, LLM providers, and API keys. Eliminates need for code changes when switching deployment modes, models, or providers — configuration is purely declarative and environment-specific.

Provides three pre-configured Docker Compose files (docker-compose-ollama.yml, docker-compose-chatgpt.yml, docker-compose-mcp.yml) that orchestrate all required services (indexer, web UI, Qdrant, LLM provider) as containers. Eliminates manual service startup and dependency management — single docker-compose up command deploys entire RAG system with correct networking and volume configuration.

Monitors local document directory for new or modified files and updates the vector database incrementally without full re-indexing. Tracks file modification timestamps and checksums to detect changes, re-embedding only affected documents while preserving existing embeddings for unchanged files. Reduces indexing time and computational cost for large document collections with frequent updates.