HR teams spend a disproportionate share of the week answering the same repetitive questions: "How many annual leave days do I have left?", "What is the notice period in my contract?", "When is the payroll cut-off this month?", "How do I claim medical reimbursement?". ERPNext already holds the underlying data in Leave Ledger Entry, Salary Slip, Employee and HR Settings, and most companies keep their rules in a handbook PDF — but ERPNext core has no conversational surface that connects the two. Employees either dig through the Frappe desk (which most non-HR staff never learn), raise a ticket, or interrupt an HR officer. ERPNext gives you the records and the workflows; it does not answer questions in plain language and it does not read your policy documents.

ECOSIRE builds an ai_hr_assistant Frappe app that adds a chat surface inside ERPNext and grounds every answer in two sources: your uploaded handbook documents and the requesting employee's own live records. On the retrieval side we ingest handbook files attached to a custom HR Knowledge Document DocType, chunk and embed them, and store the vectors so a policy question is answered from your actual text with the source section cited — not from the model's general knowledge. On the data side, whitelisted server methods (@frappe.whitelist()) expose narrow, read-only lookups such as leave balance and latest payslip, always scoped to the logged-in user via frappe.session.user, so the model never receives raw SQL access and one employee can never read another's payroll.

Technically the app is a proper Frappe module: new DocTypes (HR Knowledge Document, HR Assistant Conversation, HR Assistant Message), a Frappe REST/whitelisted API layer for the chat client script, and hooks.py doc events that re-index a handbook whenever a HR Knowledge Document is saved. A scheduler_events hourly job reconciles embeddings and prunes stale chunks. Access is enforced with ERPNext permissions and role profiles — the assistant honors the same read rights the employee already has, and every drafted HR letter or response is a review-first draft, never an auto-sent action. The LLM provider (self-hosted, OpenAI, Anthropic or Azure) is configurable in a single HR Assistant Settings single DocType, and all prompts, retrieved chunks and answers are logged for audit.

Because scope, handbook content, integrations and the chosen LLM differ for every organisation, this is a build-to-order engagement. After a short scoping call we confirm the exact DocTypes, policies, guardrails and provider, then build against your Frappe/ERPNext v15 or v16 instance. Typical delivery is one working week from confirmed scope. We build on staging, run UAT with your HR team, hand over the git repository and installable app, and provide a post-go-live support window. Nothing is generic: the assistant answers from your handbook, over your data, under your permission model.