Libra Works
In development · not yet released

LIBRA WORKS · WORK AI

Answers from your own material, finished as real output.

A work AI workspace that answers from the company documents and conversations you connect, and turns the same evidence into documents, tables and reports. Every answer shows what it read.

How we build itContact sales
The screens below are a product concept with example content.
Libra Works Permission checked

What should I look up?

Answers come from your connected material with the evidence shown, and can be turned into a document or table.

Example questions

Product concept with example data. Nothing is sent anywhere.

HOW WE BUILD IT

Four stages, each measured on its own.

Accuracy in a work assistant is not one model’s doing. Permissions, retrieval, grounding and deliverable generation are separate problems that fail in separate ways. Libra Works keeps them apart so each one can be measured.

  1. 01

    Narrow by permission first

    Filtering after retrieval hides how much you could actually read. We narrow to what you can open before retrieval runs, and resolve nested groups at query time instead of storing them flattened.

    Copied permission data starts being wrong the moment the source changes. Refresh lag is something we measure, and we re-check against the source before a passage is used as evidence.

  2. 02

    Run keyword and meaning together

    Exact strings like an employee ID or a part number are a keyword problem; typos and reworded questions are a meaning problem. We run both, merge the results, then reorder with a cross-encoder.

    Across Korean, English and Chinese material, retrieval degrades noticeably when the question and the document are in different languages. We reserve retrieval quota per language and also search with the question translated.

  3. 03

    Check each sentence against evidence

    Supporting passages are pulled before the answer is written, and any claim without one is dropped. With no evidence, Works says it does not know rather than guessing.

    Outsourcing this to a grounding check that only covers English leaves Korean and Chinese material unverified. We build it so all three languages are held to the same standard.

  4. 04

    Track deliverables cell by cell

    Tables and documents are not written as files by the model. A fixed structure is produced first and exported from there, carrying which document and page each cell came from.

    Cells without provenance stay blank and marked. Numbers are either extracted values or computed from extracted values, and the original unit and period strings are kept alongside.

This direction comes from reviewing public engineering documentation and papers. Measured figures will be published on the same basis once the product is finished.

COMPANY KNOWLEDGE

Material stays where it is. Works reads it there.

Document stores, chat and issue trackers are attached through connectors. Search follows each person’s existing permissions, so a document they cannot open is never used as evidence either.

01Documents and files

  • Document stores
  • Internal wiki
  • Shared drives

02Conversations and mail

  • Chat channels
  • Work mail
  • Meeting records (Libra Meet)

03Work systems

  • Issue trackers
  • Code repositories
  • Customer records

The list of supported services and connector types will be published at release. Below are the material types work AI generally covers.

ADMIN AND CONTROL

What an admin checks before rollout.

Work AI touches company material, so the control surface gets decided before the features do. The items below are part of the design; what is offered is settled during contracting.

  1. 01

    Permission-aware search

    Only material the person can already open is used as evidence, checked at query time.

  2. 02

    Workspace administration

    Admins set members and roles, which connectors are attached, and usage limits.

  3. 03

    Audit log

    A record of who asked what, and which material was used as evidence.

  4. 04

    Retention and deletion

    Retention periods and deletion scope for conversations and output are set by policy.

Not published yet: security certification status, data residency, and contract terms on model training. These will be published once confirmed.

WHAT IS STILL UNSOLVED

The open problems, written down too.

Some problems in this category are unsolved industry-wide. Rather than leave them out, here is how the design handles them.

Instructions planted in material

There is no known complete defence against instructions aimed at the model hidden inside a retrieved document. We treat retrieved content as untrusted input and close the route itself: the model cannot call outbound addresses or mint arbitrary links.

Misconfigured source permissions

A document shared org-wide at the source will be found org-wide. Applying permissions correctly and the source permissions being correct are two different things. Rollout includes reviewing over-shared material first.

Cross-language retrieval misses

Relevant material is missed when the question and the document are in different languages; published experiments show this clearly. Per-language quota and question translation reduce it, but do not remove it.

Rows lost from tables

Tables with merged cells can silently lose or combine rows during conversion. Row counts and totals are reconciled against the source, and mismatches are marked on the deliverable.

The public material and sources behind this are available on request.