knowledge-RAG-MCP

# knowledge-mcp MCP (Model Context Protocol) server that provides a local, file-based knowledge base with BM25 full-text search and optional hybrid (BM25 + embedding) retrieval. ## Features - **Document ingestion** — PDF, DOCX, ODT, EPUB, HTML, XLSX, PPTX, MD, TXT - **BM25 search** — Unicode-aware, CJK bigram-aware tokenizer - **Hybrid search** — BM25 + dense embedding via Reciprocal Rank Fusion (optional, requires embedding API) - **Paragraph-level chunking** with semantic merging and hierarchical (fine + coarse) sections - **Paper metadata extraction** — title, authors, abstract for academic papers ## Installation ```bash go install ./... ``` ## Usage ### Configure Set the data directory via environment variable (defaults to `<cwd>/.reasonix/knowledge/`): ```bash export KNOWLEDGE_MCP_DATA_DIR=/path/to/data ``` ### Run as MCP server (stdio transport) ```bash knowledge-mcp ``` Add to your MCP client (e.g. Claude Desktop, Reasonix): ```json { "mcpServers": { "knowledge": { "command": "/path/to/knowledge-mcp", "env": { "KNOWLEDGE_MCP_DATA_DIR": "/path/to/data" } } } } ``` ### MCP Tools | Tool | Description | |------|-------------| | `knowledge_search` | BM25/hybrid search across all documents | | `knowledge_read` | Read a specific chunk by docSlug + chunkID | | `knowledge_list` | List all uploaded documents | | `knowledge_upload` | Upload a file or batch-upload a directory | | `knowledge_remove` | Remove a document and all its chunks | ### Supported file formats `.md` `.txt` `.pdf` `.docx` `.odt` `.epub` `.html` `.xlsx` `.pptx` ## Storage layout ``` <data-dir>/ ├── INDEX.md ├── .searchlog.jsonl └── <document-slug>/ ├── meta.json ├── CHUNKS.toml ├── source.<ext> └── chunks/ ├── 000.md ├── 001.md └── ... ``` ## Dependencies - [tabula](https://github.com/tsawler/tabula) — document parsing (MIT) - [toml](https://github.com/BurntSushi/toml) — index serialization (MIT) - [mcp-go](https://github.com/mark3labs/mcp-go) — MCP protocol (MIT) ## License MIT