Content arrives two ways — add_document for text you already have in hand, or the uploads folder for files you would rather copy in — and process_uploads turns the folder into embeddings when you are ready. Retrieval is the part worth the install: search_all_documents combines full-text with vector matching for the case where you cannot say which document is relevant, and get_context_window then expands a match into the parent sections on either side, so what reaches the model is a passage rather than an orphaned chunk.
A local document store with semantic retrieval: documents arrive by call or through an uploads folder, and are embedded for search.
- Documents added with a title, content and metadata, then listed, fetched by ID or deleted
- Semantic search for chunks inside one named document, when you already know which one holds the answer
- Hybrid full-text and vector search across every document, for when you do not
- A context window of parent sections around a hit, with the number of sections before and after set per call
- An uploads folder whose absolute path and file listing you can read, and whose .txt and .md files are embedded on demand
- A web UI address for the uploads that would otherwise be a manual file copy
Disk space for the uploads folder and the embeddings built from it, and a Node runtime. No account and no key. Files placed by hand have to be .txt or .md for process_uploads to pick them up.
One command — npx -y @andrea9293/mcp-documentation-server