A handbook answer cites the section and the version

TL;DR. A knowledge pack is a recipe: keep each policy clause whole, cite the section and the version date, and re-index when the handbook changes. What we can show today is that recipe on a fictional company called Acme — a sample handbook, a library retrieval check, and a pilot that loads your file with you.

The meeting where someone says “just upload it”

People Ops hears a pitch that sounds finished. Drop the staff handbook in, and the chatbot will answer leave, benefits, and “is this allowed?”

A fluent paragraph is cheap. An answer you can paste into a staff email needs a section and a version date. The useful object is a recipe for one knowledge base: how the handbook is cut, what a citation looks like, and what you do the week the policy changes.

A recipe on Library

OrchAI Library is the memory of the suite — private Google for your team's documents, with every answer cited. An HR pack sits on that library. The product names stay Portal, Library, Agents, and Eval.

Three pieces matter in the room:

1. Cut each policy so the clause stays whole.
2. Cite the section and the version date — or the benefits page.
3. Re-index when that version changes.

We can walk those three on fictional Acme. The sample handbook is version 2026-04-01. The sample benefits guide is 2026-03-15. Acme is a demo corpus we made for this walkthrough. The sample handbook belongs to that fictional company.

Keep the clause whole

A leave rule usually holds a number and an exception in the same section. Cut the section in half and the answer can quote the number and miss the exception.

The HR recipe cuts the sample by section: one handbook section per piece, a ceiling of 400 words, and zero overlap. A benefits heading is treated the same way. The policy lives or dies as a unit.

That cut is how the sample is prepared. Until a dedicated section splitter is in place for uploaded files, an ordinary PDF can still be broken into passages. Preparing the handbook as whole sections is part of the recipe. A PDF dropped in a folder still needs that cut.

What the citation should look like

On the Acme sample, a leave answer cites the section and the version:

Handbook §3.2 (version 2026-04-01)

The question we use to show it: “How many unused annual-leave days can I carry into next year?” The sample points at §3.2, 5 days, version 2026-04-01. Those five days are Acme’s made-up rule for the walkthrough. A pilot uses the policy in the file you bring.

A benefits answer cites the provider document and the page, in this shape: Acme Benefits Medical plan p.4.

Leave and onboarding questions are meant to go through this pack and show that citation. If the answer has no section and version — or no provider document and page — hold it back from the staff email. The model points at a passage in the pack, or it stops.

When the handbook changes, refresh the pack

A version date stays honest only if the pack moves with the handbook. When the policy changes, re-ingest and re-index that knowledge base so the citation picks up the new date.

The pack reads the file you give it. Staff-directory sync sits outside this recipe. A leave or payroll spreadsheet can be read as a file in the sample shape. The sample reads that file as a document. A live payroll connection is separate work.

Say this before anyone writes a purchase line

You can say today	Still needs a separate yes
A recipe cuts a handbook by section, cites Handbook §3.2 plus a version date, and expects a re-index when the file changes.	“Your handbook is already in production.”
The sample is fictional Acme. A real pilot means we load your file and re-index it with you.	“This is a buyable HR copilot.”
On the Acme sample, a retrieval check asks the right section to appear in the top five (recall@5 at least 0.8) and the citation to match (accuracy at least 0.9).	“OrchAI Eval has signed this off.” That HR evaluation is still open. The sample check is a library test on the seed. A legal review is a separate step.
Who may open the collection is a separate switch. With Portal in the building, organisation roles stay on Portal.	“Only the HR team can see this, because the pack is labelled HR.”
Getting documents in can still mean Portal screens. The supported path still includes those screens.	“Just upload the PDF in Library and you are done.”

Where this sits

We wrote earlier about why a governed library cites the passage it recalls. A Drive stores files. A library recalls a passage and cites it. Once that library exists, an HR handbook still needs the recipe above: whole clauses, a versioned citation, and a re-index when the policy moves.

OrchAI Library holds that memory. OrchAI Agents is the front door people log into. OrchAI Eval is the quality gate when you need a scored report. Scoring this HR pack there is still open. OrchAI Portal is the full building, including who in the organisation may open what.

If the next conversation is “we were told to upload the handbook,” talk to us. Bring the version you actually use. We load that file for a pilot. The Acme sample stays a sample.

Frequently asked questions

Does the chatbot know employment law? It answers from the handbook section and version in the pack. A citation names that section and that version date. If the clause is missing from the file you loaded, the pack stops.

Is our handbook already in the product? The sample is fictional Acme (handbook version 2026-04-01, benefits guide 2026-03-15). Yours is a pilot: we load the file and re-index it with you.

What happens when the policy PDF changes next month? Re-ingest and re-index that knowledge base so the citation shows the new version date. A shared-drive edit or an HR-system edit reaches the pack when you run that refresh.

Can we buy this as an HR copilot with an Eval sign-off? You can use the Library recipe and the fictional sample. The pack is a recipe on OrchAI Library. OrchAI Eval has yet to score this pack. The 0.8 and 0.9 figures are the sample retrieval bar on that library test.