Show HN: Local-first RAG for PDF user manuals, datasheets (github.com)

A new open-source Model Context Protocol (MCP) server, askdocs-mcp, provides local-first RAG (retrieval-augmented generation) search over PDF manuals and datasheets using Ollama for embeddings and LLM inference. You point it at a docs folder with a simple TOML manifest (askdocs-mcp.toml) to register multiple PDFs; the server builds a persistent vector store (.askdocs-cache) and answers natural-language queries with page-level retrieval and citations. It’s packaged for Docker (example run command provided) and can also be run locally; configuration supports selecting embedding and LLM models (defaults shown: snowflake-arctic-embed:latest and qwen3:14b) and the Ollama URL can be overridden via ASKDOCS_OLLAMA_URL. Licensed MIT. Significance: askdocs-mcp is a privacy-friendly, offline-capable RAG stack for technical documentation—useful for embedded teams, regulated environments, or air-gapped workflows where cloud services are undesired. Technically, it leverages Ollama for both embedding and generation, auto-creates a persistent vector cache for fast repeated queries, and exposes MCP-compatible server metadata so it can be orchestrated from MCP-aware tooling. Requirements: an Ollama server with the required models pulled, Docker (optional), and a docs directory structured as shown; recommended to add .askdocs-cache to .gitignore to avoid committing the vector store.