Create and edit a Discovery without leaving chat, and keep working with finished Discoveries by attaching them as context.
Discoveries don’t have to start on the Discovery page. From any chat you can spin up a new one, watch the first cells fill in on a preview card, and, once the run has finished, keep asking follow-up questions with the Discovery attached as context.
Discovery citations resolve at four levels — the whole Discovery, a column, a row, or a single cell — and Libra is told to always cite the most specific one, so “why is row 3’s job title Innovation Manager?” lands on that cell.
You need a chat inside a project you have edit rights on: a Discovery created from chat is created in that project, and Libra builds it from all documents currently attached to the chat. Attach the documents you want as rows before you send, and remember that each row is one document — Libra reads them independently of one another, not as a single set.
Pick Tools → Create new → Discovery to commit to building one, or simply describe the extraction you want. The difference matters: with creation mode armed, Libra has to create a Discovery in that turn. Left unarmed, Libra decides for itself, and proposes a Discovery when a request amounts to asking the same questions of several documents rather than answering inline.
1
Pick Tools → Create new → Discovery
A Discovery badge appears in the composer’s toolbar row, alongside Tools and the attach button, confirming what the next message will build. Click the badge to cancel creation mode.
2
Attach the documents and describe the columns
Drop in the documents you want to extract from. Type a prompt that names the fields, like “Extract employee name, job title, and home office location for each contract”. Libra turns each field into its own column, picking the most specific answer type it can — text, yes/no, number, date, currency, or bullet list — and writes the column labels and prompts in the language you wrote in.
3
Send
Libra creates the Discovery and starts processing it straight away, then shows a Discovery created card in the conversation. The card is a deliberately small preview: at most four documents and three columns, with Columns and Documents counters above the table giving the real totals. Cells fill in on the card as the run progresses.
4
Open Discovery to see all of it
Click Open Discovery on the card to leave the preview and open the full table. Everything beyond the first four rows and three columns lives there. Use Add Column and Add Documents to extend the Discovery; new cells are queued for calculation automatically.
A Discovery created from chat is filed in the chat’s project, in the Discoveries system folder of Project files. From the full-screen view you can Save as Template — worth doing the second time you build something similar.
You don’t have to open the Discovery editor to adjust it. Attach the Discovery to the chat first — the editing tools only become available once one is attached — then ask for the change in plain language.
Add columns — “Add a column for the renewal date.”
Delete columns — “Drop column 2.”
Rewrite a column — its label, its question, or its answer type: “Make the ‘Notice period’ column answer in dates instead of free text.”
Libra applies the change straight away — there is no confirmation step — and any cells the edit creates are queued for calculation automatically, so the affected column refills without you opening the full view. Two behaviours are worth knowing: the editing tools never delete rows, and a document attached to the chat during an edit is appended to the Discovery as a new row.
Editing from chat changes the saved Discovery, not just this conversation. Switching a column’s answer type re-extracts that column, so the answers already in it are replaced.
Once a Discovery has finished running, attach it to any chat as context and ask follow-up questions. Only completed Discoveries can be read this way — one still processing won’t appear in the picker. Libra answers with citations that point at the Discovery as a whole, or at a specific column, row or cell.
With @-mention while typing
Type @ anywhere in the chat input. The picker opens in sections — Assistants, Documents, Discoveries, Reviews, Playbooks. Discoveries and Reviews list the finished items in the current project; Documents lists the project’s files and folders alongside any Databases you can search. Pick the Discovery from the Discoveries section, and it becomes part of the chat’s context.
Via Tools → Add context
Click Tools in the composer’s toolbar row and choose Add context. That opens the same picker. Same effect.
Attaching the project itself, rather than one Discovery, pulls in its ten most recently updated completed Discoveries (and its Reviews and documents) in one go. Useful when you want Libra to consider the whole matter instead of naming each artefact.
When Libra references a Discovery in its answer, the citation points at the most specific part of the table it can.
Cite level
Looks like
When
Whole Discovery
M&A Contract Review Discovery
When the answer is about the table as a whole.
Column
Notice Period
When the answer is about the whole column.
Row
Row 3, Arsenal FC Q4 Signing Agreements
When the answer is about one specific document.
Cell
Row 3, Job Title, “Innovation Manager”
When the answer is about one specific extracted value.
Hover a Discovery citation and a popover shows the neighbouring cells as a miniature table with the cited row, column or cell highlighted, so you can check the context without leaving the answer. The popover carries a Show in Discovery button; clicking either that or the citation itself opens the Discovery and pops the cited row open at the cited column.
Discoveries are project-scoped, and the picker only lists the active project’s. The chat needs to be in the same project — or attach the project itself.
2
Type @ and pick the Discovery
Completed Discoveries appear under Discoveries, sorted by title, each with the Discovery table icon.
3
Ask your question
Anything from filter to aggregation to a free-form summary.
4
Read the answer with its citations
Inline citations link into the table. Open the Cited Sources panel and the Discovery gets a card with the same miniature table preview and an Open Discovery button.
Behind these, Libra either reads a range of rows and columns straight out of the table or hands the question to a data-analysis sub-agent that works over the table as a CSV — which is why aggregations and rankings hold up rather than being eyeballed.
Filter the table
“Which rows have a notice period longer than 90 days?”Libra returns a filtered subset, with citations to the specific cells.
Sort or rank
“Rank the contracts by contract value, descending. Show the top 10.”
Aggregate
“What’s the average notice period across all rows?” / “How many contracts have a non-compete clause?”
Compare
“How does the indemnification scope differ between rows 3, 5, and 7?”
Draft summaries
“Draft an email summarising the highest-value contracts for the partner. Include counterparties and contract values.”
Translate
“Translate the high-value rows’ counterparty names and contract values into German.”
Attach the documents before sending (creation flow)
The Discovery is built from whatever is attached at the moment you send, and nothing else. Attach a document later and it becomes a row only when it is added — with Add Documents in the full view, or automatically when you attach it in the same turn you ask Libra for an edit.
Be specific about what you want
“Summarise this Discovery” gets generic. “Summarise the high-value contracts in row 1, 4, and 7, emphasising non-compete duration” gets something usable.
Ask for citations back
“And cite the relevant cells” on any prompt gets you a verifiable answer.
Stack context if you need it
Attach the Discovery and a Review and the underlying documents. Libra can cite into all of them; “explain why Topic 3 in the Review is medium risk and which Discovery rows are most affected” is a perfectly usable prompt.
Don't re-extract data in chat
If you find yourself asking “now extract the notice period for each contract” in a chat, add a column to the Discovery instead. A column runs the same question over every document and gives you a cell you can cite; a chat answer gives you prose you have to re-check.