Your documents, the index and the model that answers from them all sit locally, which is a different trade from any hosted retrieval connector — and the reason to pick this one. You pay for it up front, in an Ollama install and a model download before the first question returns anything.
A local retrieval store an assistant can both add to and read from. Documents are parsed with Unstructured so titles, lists and tables survive instead of being flattened into a wall of text, with fallback strategies underneath so an awkward file still gets in. Generation runs through a model you serve locally with Ollama and embeddings are computed on your machine, so once the models are pulled nothing leaves it.
- Facts, definitions and notes stored under a source name you choose, so they can be asked about sessions later
- File ingestion across PDF, Word, PowerPoint, Excel, OpenDocument, HTML, XML, Markdown, plain text, CSV, JSON, YAML, email, and images through OCR
- Answers that come back with the sources they were drawn from and how relevant each one was
- Filtered search that narrows by file type or by structure — only documents containing tables, or at least so many titles
- A report of what the knowledge base currently holds, broken down by file type, structure and how each document was processed
- A desktop application for bulk-ingesting a folder, with a preview-and-select step before anything is written to the store
- A Markdown copy of every processed document, kept on disk for checking what was actually indexed
Python 3.10 or newer, and Ollama installed and running with a language model pulled before the first question will work — that is the step the system genuinely depends on, not an optional extra. Tesseract is needed only for reading text out of images. The project is built around Windows: installation, launch and diagnostics are .bat scripts and the documented client entry points at one of them, so on macOS or Linux you are running the Python underneath by hand. Its documentation is in Spanish. The vector store and the converted document copies live in folders inside the project.
Build from source — clone the repository and build it, then point your client at the binary