LOCAL RAG OVER YOUR OWN DOCUMENTS
Local retrieval over your own documents, as a finished product
You know the pattern: chunk the documents, embed them, store the vectors, retrieve on a question, ground the generation, cite what was used. GI JI is that, shipped as an application you install rather than a stack you assemble. This page is written so you can decide against it quickly if it is the wrong shape for you, because for a good number of people reading this, it is.
Where building your own is the better answer — said first, on purpose
If you want to swap the reranker, tune chunking, or measure a pipeline change against your own evaluation set, build your own. GI JI lets you choose models by role and manage them from the Models card; it does not expose chunking or retrieval-pipeline tuning. If you need to read the source, audit it, or fork it, build your own: GI JI is not open source. If you need OCR, or a team sharing one index, or a service another program calls, build your own. And if the assembly is the part you enjoy, that is a perfectly good reason on its own.
What GI JI actually is, in mechanism terms
A desktop application with a local web interface. It serves its own screens over loopback. The configuration check rejects any listen address that is not 127.0.0.1, ::1 or localhost, and the server refuses to start rather than binding somewhere another machine could reach it. You cannot turn it into a shared server, and that is the point rather than an oversight.
It manages local models by role. The installer reuses a compatible Ollama installation or fetches a private runtime when needed. GI JI fetches its initial models, about 4 GB, the first time you open it, and the Models card lets you load, unload, get, or remove role choices. You need roughly 6 GiB of free disk and an internet connection for setup. After setup, GI JI needs no internet to read your documents and answer questions from them.
The index is yours and it is local. Your documents stay where you keep them. The index, questions, saved notes, and backups GI JI builds live in the GI JI folder on your own disk. There is no GI JI cloud library to sign into. Licence activation contacts Freemius with what is needed to verify the licence.
Retrieval is bounded by an explicit mode. Sources only answers from your imported material and stops when it cannot attach enough usable evidence — no answer-writing call is made. Guided adds Working Knowledge you have reviewed and approved. Open may add clearly marked general context, which is never cited to your documents. The boundary is a visible control, not a hidden prompt setting.
Connected writing is optional. Setting up a key contacts OpenAI to list the models available to that account. Connected AI is off until you switch it on. Before it writes anything for you, GI JI stops, shows you what it wants to send, and waits for you to approve that one request. You can keep the API key for the current session or choose to remember it on this computer. A remembered key is stored in data/connected.json inside the GI JI folder, excluded from GI JI backups, and removed when you choose Forget this key.
The input formats, exactly
Text-based .pdf, .docx, .txt, .md, .html and .htm, folders of them, and permitted websites. PDF text extraction goes through Poppler's pdftotext. There is no OCR: a scan-only PDF has no text layer to index, so GI JI refuses it rather than importing an empty document and pretending it worked.
The platform truth
GI JI 0.13.0 is a production release for macOS, Windows and Linux. Windows is available from the Microsoft Store or as a direct download.
The one-line version
If you want control over the retrieval pipeline, build your own — genuinely, and this page has tried to make that easy to decide. If you want a local document assistant with inspectable citations and a Sources only mode that can stop when support is missing, that is GI JI, for $79 once.
WATCH IT WORK
Answer, cited passage, and an honest refusal.
Ask a question of your own documents, open the cited passage, then use Sources only for a question the library does not support and watch GI JI stop without writing an answer.
HOW GI JI COMPARES
Where GI JI wins, and where it does not.
Three comparisons with the tools you are most likely choosing between, each one naming the case where the other tool is the better buy.
WRITTEN FOR ONE READER
Four pages that answer one question properly instead of all of them briefly — what a citation is actually worth, what GI JI is and is not for a technical reader, how it stands beside a note library, and what it does for work whose material cannot leave the computer.