Knowledge Base & Indexing

Give SphereHarness your own documents to answer from, index your parts so you can find them by describing their shape, and store the facts it should remember about how you work.

Three separate stores

  • Documents — PDFs, Word files, spreadsheets, Markdown, plain text. Searchable, and quoted back when relevant.
  • Parts — a searchable summary of a CAD part's geometry, so you can find it later by describing what it looks like.
  • Memory — short durable facts about you and your work: units, materials, current project.

Indexing a document

Use the Index Document button (the book icon in the chat toolbar), or just ask. Accepted formats are PDF, DOCX, XLSX, MD and TXT.

Index C:\standards\drawing_standard.pdf into the knowledge base

A document can be filed as a general document or as a spec. Specs are the ones you want treated as authoritative — material standards, fastener tables, company drafting rules.

Index C:\standards\ISO_2768_extract.pdf as a spec called "General tolerances"

How it works

The file's text is split into small chunks; each chunk is converted into an embedding — a numeric representation of its meaning — and stored. When you later ask a question, SphereHarness retrieves the passages closest in meaning to your question and answers from them. That is why a question phrased in your own words still finds the right paragraph, even when it shares no keywords with it.

Asking questions of your documents

What does our drawing standard say about title block placement?
Search the specs for the general tolerance on a 120 mm untoleranced dimension
According to our fastener table, what clearance hole should I use for an M8 socket cap screw?

You can also put an indexed answer straight to work:

Look up our standard clearance for M8 in the knowledge base, then add four holes at that diameter to this plate, 20 mm in from each corner
Retrieval is not comprehension. SphereHarness answers from the passages it retrieved, and it can retrieve the wrong ones — particularly when a standard has near-identical tables for different cases. For anything that matters, ask which document and section the answer came from, and check it.
Which document and section did that tolerance come from?

Where your documents are indexed

A document you index goes to a document index — a small search service — and never to the SphereHarness server. Choose it in the desktop agent: Settings → Knowledge → Document index.

  • This PC (the default) — the “Document indexing service” installed with your add-in, in your own profile. Nothing leaves the machine.
  • My organisation's document index — the same service run by your company on its own server, so colleagues search the same documents. You sign in with your Windows login (no key), with Microsoft Entra ID (Sign in with Microsoft, once), or with a shared key kept encrypted on your PC. The index decides by group who may search and who may upload or delete.

Your organisation can set this for everybody — in the org portal, or by Group Policy on each PC — and the fields are then shown locked. Test says what answered and who you are signed in as. The SphereHarness server is not offered as a document index; a PC that had chosen it indexes on this PC and says so. IT: Document Index: IT Setup.

Indexing parts — search by shape

“We made something like this two years ago” is a real and expensive problem. Indexing a part stores a summary of its geometry so you can find it again by describing it.

Index this part into the part knowledge base with part number BR-2041

The extraction happens locally on your workstation; only the summary is uploaded. Later, from any document:

Find an indexed part that is a U-shaped clamp about 80 mm wide with two through holes
Find similar indexed Inventor parts to the one I have open

You get back candidates with their part numbers and file locations; open the one you want and carry on from it.

Memory — facts SphereHarness keeps

Memory on the SphereHarness server is not offered. It was kept on our servers, and SphereHarness no longer keeps anything from your conversations there, so the remember and recall requests below are refused. A coding agent may keep its own memory, on your PC or with its provider; for house rules, use a skill, which is a file on your PC. The rest of this section applies only to an organisation that runs its own SphereHarness server with hosted services switched on.

Memory is per-user and persists between conversations and between sessions. Use it for the standing facts you would otherwise repeat every day.

Remember that our default material is 6061-T6 aluminium
Remember that our default general tolerance is ±0.1 mm and that I always work in millimetres
Remember that I am currently working on project FRAME-2026 in C:\projects\frame

Later, in a new conversation:

What tolerance do we use by default?
What do you remember about my current project?

And when something changes:

Forget what you remembered about the default material — it is now 5083-H111

Memory versus a skill

They overlap, and the distinction is worth getting right:

  • Memory holds facts — “our default material is 6061-T6”. Private to you.
  • A skill holds instructions — “always assign the default material to a new part without asking”. Shareable across the organization.

Facts you would tell a colleague go in memory. Rules you would put in a work instruction go in a skill.

Past conversations

Your own chat history is searchable too, which is useful when you know you solved something before but not how.

Search my past conversations for how we set up the bracket stress study
Chat history is memory, not truth. It records what was said, including anything that was wrong at the time. Treat a recovered answer as a lead to re-verify, not as a result.

Managing what is stored

Indexed documents are listed in the desktop agent's Settings → Knowledge → Indexed documents, where you can add and remove them. They stay in the index you chose — on your PC, or on your organisation's own server — and are not used to train models.

See Organization Admin Guide for the portal pages, and for who in your organization can see what.

Getting good answers

  1. Index the source, not a summary of it. The original standard retrieves better than someone's notes about it.
  2. Split enormous documents by subject where you can. A 600-page manual indexed as one file competes with itself for relevance.
  3. File authoritative sources as specs so they are distinguishable from general reference.
  4. Ask for the source whenever the answer will drive a design decision.
  5. Re-index when a document is revised. Nothing detects a superseded standard for you.

See also