Build org knowledge the copilot can cite

Overview

Org Knowledge is the organisation's memory for the copilot, and the controls over its second, live source. Everything the copilot says comes from one of four places: the metadata MetaLens360 has synced and indexed, the documents your team uploads here, the learnings your team has verified, and — when a connector is set up — the org itself, read live at answer time. The page shows all of them and who may see or use each.

• Sources — uploaded documents and pasted notes (KT packs, runbooks, incident reports, architecture decisions). Each is split into passages and becomes searchable within seconds of upload.
• Learnings — a problem and its resolution, captured from a copilot answer or written by hand. A learning starts as a candidate and is shown to colleagues only once a curator verifies it.
• Domains — explicit groups of orgs that may share knowledge with each other, for example the ten orgs of one business unit. Orgs that are not in a common domain never see each other's knowledge.
• Evaluation — a repeatable benchmark run against your own data: exact-count accuracy, workspace isolation, retrieval, citation coverage and index freshness.
• Live data — the MCP connectors the copilot may read an org through right now (a Salesforce hosted MCP server or any Streamable HTTP MCP server), a guarded Try a SOQL box, and the switches that keep live access read-only and audited. See Read live Salesforce data through MCP.

Every answer in the copilot shows a Grounded / Partly grounded / Low confidence label and an evidence list with markers such as [S1] (document), [L1] (learning), [F1] (org index), [C1] (component) and [V1] (a live read, with the connector and the time), so you can see what it rested on. An Indexed or Live chip under each answer says which source it came from.

Before you start

• Reading verified knowledge needs no special permission. Adding documents and proposing learnings needs Contribute org knowledge; verifying, retiring and defining domains needs Curate org knowledge. Admins and Architects have both by default. Adding and testing live connectors needs Manage live data connectors (Admin, Architect, DevOps); letting the copilot read live data on your behalf needs Query live org data (every role except Viewer). See Roles and permissions.
• Supported uploads: Markdown, text, CSV, JSON, HTML, Word (.docx) and Excel (.xlsx, .xlsm), up to 15 MB. PDFs are not read yet — export them to Word or text first. Of any document, the first 2,000,000 characters of text are kept and indexed.
• Word and Excel files are compressed, so they have a second limit once unpacked: a Word document's text may be at most 50 MB, and the parts of an Excel workbook that hold its cells, sheet names, text and number formats at most 100 MB in total (pictures and other embedded files are not unpacked and do not count), with at most 256 sheets. A file over a limit is refused with a message that says which, for example This workbook is larger than the 100 MB that can be read once uncompressed. Split the file, or save it as text and upload that. or This workbook has more than 256 sheets, which is beyond what can be indexed. Split it into smaller workbooks.
• A workbook within those limits is indexed sheet by sheet, up to 3000 rows a sheet, with a few further caps that end in a note rather than a refusal: only the first 1000 columns of each sheet ((Columns after column 1000 of this sheet were not indexed.)), the first 1,000,000 distinct text values and at most 12,000,000 characters of the workbook's shared cell text — cells past either are left empty, with the note (Some cells were not indexed: this workbook has more than 1000000 distinct text values, or more shared text than can be indexed at once.) — 32,767 characters per cell, and a reading budget — at most 5,000,000 characters of text and a bounded amount of workbook structure — after which the text ends with (The rest of this workbook was not indexed: it is larger than what can be indexed at once.); the 2,000,000-character limit above still applies. Error cells show their error text, such as #N/A, and a merged cell's value appears once, in its top-left cell.
• These Word and Excel files are refused as well, each with a message saying what to do: password-protected files and old .doc / .xls files renamed to .docx / .xlsx (This file is password-protected or in the old .doc/.xls format, which cannot be read. Remove the password (or save it as .docx/.xlsx) and upload again.), damaged or truncated files, files with extra data in front of their contents (This file is not a plain Office document (it has extra data in front of its contents), so it was refused. Re-save it from Word or Excel and upload again.), files in a format Word and Excel no longer write by default (ZIP64, an unsupported compression method — re-save them from Word or Excel), and files with an unusually large inner structure (more than 5000 internal parts). Nothing of a refused file is indexed.
• Treat uploaded documents as data, not instructions: the copilot is told to never follow text found inside a document.

Steps

Upload a document

1. Open Org Knowledge from the left navigation (under Overview) and choose the org in the top-right picker.
2. On the Sources tab click Upload a document (or Add a note to paste text).
3. Choose the file, give it a title and a kind (document, runbook, incident, KT, decision…).
4. Choose who may see it: one org, a knowledge domain, or the whole workspace. A platform admin choosing The whole workspace also picks which Workspace it is for; everyone else's workspace-wide notes go to their own workspace.
5. Optionally set Treat as stale after (days). Past that age the copilot still uses the document but says it may be out of date.
6. Click Index now. The document is listed with its passage count and can be cited immediately.

Uploading a newer copy later? Upload it with the same title and the old version is retired automatically when the content differs.

Capture a learning from the copilot

1. In the copilot, under a helpful answer, click the bookmark icon (Save as a reusable learning).
2. A draft is written from the question and the answer. Check every line — it is a draft, not a fact.
3. Choose the scope: This org only, A knowledge domain, or Only me.
4. Click Save candidate. A curator can tick Mark as verified now instead.

You can also click the thumbs-up or thumbs-down under any answer; the feedback is recorded against that answer for the evaluation.

Review candidates

1. On the Learnings tab, open To review.
2. Open a candidate. If it overlaps a verified learning, the overlap is shown — tick the one it replaces to mark the older one superseded.
3. Add a note and click Verify, or Reject.

Share knowledge between orgs

1. On the Domains tab click New domain, name it, and tick the member orgs.
2. Scope documents and learnings to the domain when you add them. When the copilot uses one for a different member org it says Shared via domain — confirm it applies here.

Check retrieval and run the evaluation

• On Sources, type a question under Test retrieval to see exactly which passages and learnings the copilot would be given for the selected org.
• On Evaluation, click Run evaluation. Tick Include AI citation checks to also spend a few provider calls measuring whether verified learnings are cited when their problem is asked about.

Connect live org data

1. On the Live data tab, read the chip next to the org name: it says whether live data is available for the selected org and, if not, why.
2. Click Add connector, fill in the endpoint and credential, save, then click Test on its row. The tools the server offers are listed with Read-only / Write badges, and the status turns Connected.
3. In the copilot, leave the toggle on Auto (record questions go live, metadata questions to the index) or pick Live.

Connectors are read-only by default. The Writes switch only lets the copilot propose a tool that changes data; each proposal still waits for Approve and run in the chat by a holder of Manage live data connectors, and every call is audited. The full setup, including the Salesforce side, is in Read live Salesforce data through MCP.

Example

Priya resolves a CPQ pricing failure in acme-uat after a long chat with the copilot. She clicks the bookmark under the final answer, keeps the drafted title QCP timeout on quotes with more than 200 lines, chooses the CDW domain and saves. Arjun, a curator, sees the candidate under To review, notices it overlaps an older learning that recommended a different fix, ticks it as superseded and verifies. The next day Sam, working in acme-prod (also in the CDW domain), asks why quote calculations fail and the answer cites Priya's learning as [L1] with the note Shared via domain "CDW".

FAQs

Does a candidate ever influence other people's answers?

No. Only verified learnings are retrieved for colleagues. A candidate is shown back to the person who proposed it, marked as their own unreviewed candidate.

Can people in another workspace see our documents?

No. Knowledge never crosses a workspace. Inside a workspace, org-scoped knowledge is retrieved only for that org or for the orgs of a domain it is shared with.

What happens when I retire a document?

It is removed from every future answer immediately. Learnings that cited it keep working but are flagged (source retired) and their confidence is lowered, so a curator can decide whether they still stand.

What does "used N×" on a learning count?

How many times the learning was retrieved to answer something: a copilot or chat answer, the knowledge search, or a live meeting suggestion. Comparisons do not count — the Meeting Copilot checking whether an extracted item is new, and the review's list of similar items, leave the counter alone.

Does this retrain the AI model?

No. New documents and verified learnings are indexed and retrieved; no model is trained or changed. That is why they are usable within seconds.

Does a live answer become part of the knowledge base?

No. A live read is a snapshot, labelled with its time, and is not stored as knowledge. If the method was useful — the SOQL, the tool, the interpretation — save the answer as a learning; the number itself is re-read the next time someone asks.

Troubleshooting

"Identical content is already indexed"

The same text was uploaded before under another title. Open that document instead, or change the content.

"No AI provider is available to compose an answer"

The copilot still returns the evidence it found (learnings, documents, org facts) as a list. Add or fix a provider under Settings → AI Providers. See Configure AI providers.

A count looks wrong

Counts come from the org's last sync. Check the org's sync date in the copilot's org strip or on Org Sync & Repo Fetch Status, and sync again if it is old. A type at zero may simply not be in the org's sync selection.

Related articles

• Use the AI assistant
• Read live Salesforce data through MCP
• Configure AI providers
• Roles and permissions
• Org Overview
• Access 360