No documents indexed yet
Add a folder of documents and emails, and they become searchable in
seconds. Supported: .docx .doc
.pdf .msg .eml
Start a draft
Describe what you need and the closest precedent in your documents is found and drafted from, or upload a template and get the same document with this matter's details in it. Either way nothing is invented: anything without a source comes back flagged for you.
Folders
Your documents keep the folder structure they arrived in. Click a folder to see what is inside, or to search only within it.
Remove documents from the library
Clears one folder from the library so it can be added again from scratch. Use it when documents were filed under the wrong folder, or were read badly. Updating again will not fix those on its own — a file that has not changed is skipped, so it is never re-read, even when the problem is in how it was read the first time. Removing it here is what makes an update treat it as new.
Documents
Passages
Add documents
Supported formats: loading…
Drop a folder or files here
Sub-folders are kept, so a matter stays a matter.
Without this, a file already in the library is skipped if its content is the same, and reported “unchanged”, however it was read the first time. Tick it to have these files read from scratch — scanned pages go through character recognition again, so it is slower.
Or update from a folder already on the server
Anything placed in the mounted documents folder can be added without uploading. Unchanged files are skipped, so running this again is quick — and it picks up anything that could not be read last time.
Every file is read from scratch and scanned pages go through character recognition again, so expect this to take as long as the first run did. Only needed when a document was added but read badly.
A document is filed under its path relative to the folder named here, so updating from a sub-folder re-files its documents under a different matter. Leave this empty unless you know that is what you want.
Recent activity
Accuracy check
Runs a fixed set of test questions through the real pipeline and scores the results, so accuracy is measured rather than assumed.
History
Benchmark: models compared
Real patent documents and questions, scored against an answer key, on the
model the application is using now. Search benchmark asks your own search
the questions on documents already in the Library - once scoped to the right
document (is the answer right?), once across the whole Library (is the right
document found?). Upload benchmark tests the attach-a-file features and
needs the documents' text. To compare several models at once, use
scripts/run_benchmark.py.
Feedback
Accounts
Everyone who can sign in. Passwords are stored only as hashes, so an existing password cannot be shown - to change one, set a new one. There are no roles: every account can do everything, including managing these accounts.
AI model providers
Two places call a model: answering a question, and writing a draft.
Each is set independently in .env and can use a different
provider. Both have a default that needs no API key at all.
Settings in effect
Every value comes from the .env file. Change it there and
restart to take effect.