Key facts
| Pipeline | Handbook ingestion → chunking → embeddings → cited answers |
| Embeddings | plugsky-embed and plugsky-embed-multilingual via /v1/embeddings (live) |
| Generation | Cited answers with JSON mode on /v1/chat/completions (live) |
| Metadata | Country, policy version, effective date and section per chunk |
| Models | 30+ models, including multilingual options for mixed workforces |
| Access | Scoped API keys and RBAC; retrieval filters by workforce segment |
| Residency | Region-locked planes, VPC, on-prem and air-gapped options |
| Free tier | plugsky-micro and plugsky-lite on the free plan, no card required |
TL;DR
- Ground every answer in the handbook clause that governs it — and show the citation.
- Store country, effective date and version metadata so local rules win correctly.
- Answer from policy text, never from the model's assumptions about employment law.
- Keep employee data out of prompts entirely; personalise with a follow-up step if needed.
- Flag stale or conflicting clauses instead of guessing between them.
How it works, step by step
- Assemble the corpus: global handbook, country addenda, benefits guides, leave policies and process documents, each with owner and effective date.
- Chunk by clause and section, keeping headings, country and version in metadata so retrieval can filter.
- Embed chunks with plugsky-embed, or plugsky-embed-multilingual when policies and queries span languages.
- Build a retrieval step that resolves the employee's country and workforce segment, then applies those filters before ranking.
- Generate the answer with an instruction to quote or cite the governing clause and to flag conflicts between documents.
- Return structured output with the answer, the citations and a fallback link to the source policy page.
- Test with questions across countries, including questions where policies conflict, and review with HR before launch.
Original data
Try it yourself
Open the RAG chunk size calculator →
Grounding HR answers in policy documents
The value of RAG in HR is not fluency — it is provenance. Employees and managers need the answer that their jurisdiction's policy actually gives, with a citation they can open. Retrieval supplies the governing text; generation explains it in plain language; the citation makes the answer checkable by HR and by the employee.
The model calls are live on Plugsky: /v1/embeddings for indexing and /v1/chat/completions for grounded answers with JSON mode. Documents, filters and permissions remain in your infrastructure, which is where employment data belongs.
Chunking handbooks, contracts and country addenda
HR corpora are hierarchical and jurisdictional, so generic chunking fails. A global policy with per-country deviations must keep those deviations attached to the right scope. Chunk by clause and carry the scope in metadata: country, region, workforce type, effective date and version.
- Clauses over pages: one chunk per numbered clause or policy section, with its heading.
- Addenda first: when a country addendum modifies a global clause, retrieve both and let the answer state the local rule.
- Effective dates: store them and prefer the current version; surface upcoming changes separately.
- Keywords too: keep lexical search for exact terms such as benefit names and form numbers.
Privacy, access and evaluation
The cleanest privacy design is to never put employee data into the retrieval prompt. Answer policy questions from documents, and if personalisation is required, run a separate step that checks entitlements in your systems and merges the result. Retention settings should match HR policy, and deployments can be pinned to a region or run on-prem where rules demand it.
Evaluate with questions that have verifiable answers per country: does the response cite the correct clause, does it prefer the local addendum, and does it refuse when policies conflict or the corpus is silent. Track escalation to HR and keep a review queue for flagged conflicts. Refresh the index whenever policies change, and record the version that produced each answer for audit.
Honest comparison
| Capability | Plugsky RAG | Static FAQ page | General chatbot |
|---|---|---|---|
| Answer basis | Retrieved policy clauses with citations | Curated entries | Model assumptions |
| Jurisdiction | Country and version filters in metadata | Manual sections | Unreliable |
| Freshness | Re-index on policy change | Manual edits | Training cutoff |
| Data exposure | Documents only; no employee data needed | None | Depends on provider |
| Residency | Region-locked planes, VPC, on-prem, air-gapped | Static hosting | Vendor-controlled |
Frequently asked questions
Does the assistant need employee data?
No. Answer from policy documents. If you need personalised entitlements, resolve them in a separate tool call and merge the result without sending personal data to the model.
How do we handle country-specific rules?
Store country and region in chunk metadata, filter retrieval to the employee's jurisdiction, and instruct the answer to prefer the local addendum and cite both documents.
Which embedding model handles multilingual policies?
plugsky-embed-multilingual embeds queries and documents from multiple languages into a shared space, so a question in Arabic can retrieve an English policy clause.
Can we keep everything inside our network?
Yes. Plugsky supports region-locked cloud planes plus VPC, on-prem and air-gapped deployment for policy corpora that cannot leave your environment.
How do we stop answers from sounding like legal advice?
Constrain the assistant to policy text, require citations, and add a standing disclaimer that it explains policy rather than providing legal advice.
What happens when policies conflict?
Instruct the model to surface the conflict and cite both clauses rather than choose silently, and route flagged cases to HR for resolution.
How often should we re-index?
Re-index whenever a policy, addendum or benefits guide changes. Store effective dates and versions so superseded text can be filtered or shown historically.