That is the trade: a heavier install — Git LFS, a local embedding model, a first-run embedding pass the README puts at roughly 20 minutes — in exchange for search that costs nothing per query and sends nothing anywhere. Budget the setup time, and remember to run `git lfs pull`; without it the clone succeeds and the library is empty.
A self-contained knowledge base of consulting and industry reports, packaged as an MCP server. The reports ship with the repository and are stored locally; retrieval works both by keyword and by vector similarity against a bundled embedding model, so nothing has to leave the machine to answer a question.
- A bundled corpus — the shipped statistics record 174 files across 9,320 pages, 9 publishers including McKinsey, PwC, BCG, Bain, Deloitte and Accenture, and 93 topics
- `get_similar_content_by_rag` finds the closest passages by vector similarity, using the built-in `Qwen3-Embedding-0.6B` model
- Keyword retrieval over report details for when you know the term you want
- Pagination on results, and the local path of each referenced file so you can open the source
- Your own PDFs join the same library: drop them in `library_files`, run `main.py`, and they are extracted, parsed and optionally embedded
- Report pages can be browsed online from inside the MCP client
Python 3.12 or newer, plus Git and Git LFS — the corpus comes down through `git lfs pull`, so a plain clone leaves you with an empty library. Published as InsightsLibrary version 0.1.0. Setup is `uv venv .venv`, activate, `uv install .`, then register `uv` with `--directory` pointing at the project root, followed by `run` and `ikb_mcp_server.py`. Nothing is needed for the bundled reports. Adding your own PDFs is the only part that needs credentials: set `VLM_API_KEY`, `VLM_BASE_URL` and `VLM_MODEL_NAME` in `.env` for the vision model that reads them.