rag-memory-epf-mcp

Implements a retrieval-augmented generation system that stores and indexes project-specific documents locally using vector embeddings, enabling semantic search across a knowledge base without external cloud dependencies. The system maintains embeddings in a local vector store and performs similarity-based retrieval to augment LLM context with relevant project information, supporting multilingual content through language-agnostic embedding models.

Builds a graph-based representation of relationships between documents, entities, and concepts extracted from project knowledge, enabling structured reasoning and multi-hop retrieval across connected information. The system likely uses entity extraction and relationship inference to construct nodes and edges, allowing agents to traverse semantic connections rather than relying solely on vector similarity.

Implements semantic search across documents in multiple languages using embeddings that map different languages to a shared vector space, enabling cross-lingual retrieval without language-specific models or translation preprocessing. The system likely uses multilingual embedding models (e.g., multilingual-e5, LaBSE) that natively support 50+ languages, allowing a query in one language to retrieve relevant documents in any language.

Exposes RAG and knowledge graph capabilities through the Model Context Protocol (MCP), allowing Claude and other LLM clients to invoke memory operations as tools within agent workflows. The server implements MCP's resource and tool interfaces, enabling agents to call memory retrieval, graph traversal, and search operations as first-class capabilities without custom integration code.

Processes raw documents (markdown, code, text) into indexed vectors and knowledge graph nodes through a pipeline that handles chunking, embedding generation, and metadata extraction. The system likely implements configurable chunking strategies (sliding window, semantic boundaries) and batch embedding to efficiently process large document collections while maintaining chunk-to-source traceability.

Splits documents into chunks optimized for semantic coherence rather than fixed-size windows, preserving context boundaries to ensure each chunk contains complete concepts. The system likely uses sentence/paragraph boundaries, code block detection, or semantic similarity thresholds to determine chunk boundaries, maintaining references to parent documents and surrounding context.

Enhances search queries by generating related terms, reformulations, or sub-queries to improve retrieval coverage, using techniques like synonym expansion, query decomposition, or multi-query generation. The system may use LLM-based query expansion to generate semantically similar queries that retrieve documents missed by the original query, or decompose complex queries into simpler sub-queries for targeted retrieval.

Enables filtering search results by document metadata (type, source, date, tags, language) and supports faceted navigation to narrow results by multiple dimensions simultaneously. The system maintains metadata indexes alongside vector indexes, allowing hybrid queries that combine semantic similarity with structured filtering, enabling agents to constrain searches to specific document types or sources.

Intelligently selects and ranks retrieved chunks to maximize relevance within LLM token limits, using techniques like diversity-aware ranking, importance scoring, and redundancy elimination. The system may re-rank results by relevance, remove duplicate information, and prioritize high-impact chunks to fit within the LLM's context window while preserving the most important information.