Instead of toggling between three windows — reports, patents, papers — you ask from one channel, and every answer arrives with a document name, a page, metadata

Three kinds of evidence, scattered across three systems

In pharma and chemical R&D, a single judgment rarely stands on material from one place alone. Internal technical reports accumulated over time, patents that draw the lines of rights, the outside world's latest academic papers. Usually you have to look at all three at once. The problem is that these three mostly live in different systems and repositories.

• Internal technical reports. They pile up in the report system, and you find past documents by rummaging through them by keyword
• Patents. You find them by setting conditions in the patent search tool. The claims draw the scope of rights
• External academic papers. They live in an academic DB and carry metadata like title, author, source, and publication info
• Each system's search method and screen is different. To gather the basis for one question you have to toggle between systems several times, and cross-checking the pieces is the researcher's own job
• Internal resources alone make it hard to keep up with the fast-moving flow of the latest research. You want to see inside and outside in one flow, but the tools run separately

Four searches, one manual stitch

The old approach runs in roughly four steps. You rummage through past documents by keyword in the internal report system. You set conditions in the patent search tool to find related patents. You search papers in an external academic DB. Finally, a person directly cross-checks and organizes the evidence gathered from the three places.

This flow has two fractures. One is the keyword trap. If the search term diverges even slightly from the field's domain terminology, the very evidence that matters drops out of the results. Even for the same substance or the same process, if a report and a paper use different expressions, keyword search can't fill the gap between them. The other is the separation of sources. Since you have to note the answer and the source separately, when you later write a report or cite a conclusion, you wander around looking for where that was again.

I found the answer, but then I spend more time finding where that answer came from.

Put all three lines into one workspace

The setup is getting all three kinds of material into one workspace. Reports and patent documents go in as files, external papers as links. Your originals stay as they are, because the app reads a copy.

Ask all three lines from one channel

"What do the internal reports have on this substance?"
→ The agent searches the internal technical reports and answers.
   The answer cites "which document, which page" alongside.
   Conclusion and source come out together.

"Which patents catch this composition range and these process conditions?"
→ Not by simple keywords, but by real research conditions like
   composition ranges and process conditions, it filters relevant patents.
   It becomes the starting point for reviewing the scope of rights.

"How far has the outside research on this topic come?"
→ It researches papers based on title, author, source, and
   publication-info metadata. The internal basis and the outside
   world's latest research connect within one flow.

(after the three lines are gathered)
→ The gathered evidence and sources come back organized as a report.
   You start the review from already-organized results, not a blank screen.

Sources ride along with the answers because the graph has already read the documents. On an entity page, hover a connection chip and the source sentence behind the fact appears, cited as "{note} · line {n}". With conclusion and source arriving together, there's less need to note the source separately.

It doesn't throw all three lines at once with no order. It sets the priority of what to look at first, and handles repeated lookups fast by storing a result once it's fetched. It doesn't entrust everything to a single model either. Multiple LLMs (large language models) are being comparatively evaluated to home in on the configuration that best fits this work.

What changes

What the PoC sets out to validate is this structure: gather three scattered kinds of evidence into one flow, and accompany every answer with its source. How much that structure raises the speed and trust of actual research judgment hasn't been measured yet — that's the stage the PoC is validating.