HRagentAssistant

Employees ask HR the same policy questions repeatedly, while a generic chatbot can invent answers, ignore document versions or expose restricted documents.

A document-grounded assistant: it retrieves passages from the HR documents actually loaded, answers with verified citations, abstains when the evidence is not enough, reports conflicts between document versions and enforces role-based document access in the backend. It also guides a leave request (summary, explicit confirmation, HR decision) and lets an employee forward an unanswered question to HR.

Personal project built as a local prototype on synthetic data: retrieval pipeline, guardrails, authorization model, workflows, Streamlit interface, test suite and documentation (architecture, security, roadmap).

• Answers carry citations that are checked against the retrieved passages; the assistant abstains instead of guessing
• Detects conflicting active versions of the same document and reports the conflict
• Role-based access (employee, HR manager) enforced in the retrieval filter and the backend, not in the prompt
• Document content is treated as untrusted: prompt-injection attempts are neutralised
• Works fully offline by default (extractive answers, local embeddings, embedded Qdrant); an LLM and semantic embeddings are optional
• Leave-request workflow with explicit confirmation and HR approval, and escalation of questions to HR (no notification is sent)

• Abstain rather than answer: a lexical coverage gate refuses questions the documents do not cover.
• Access control is applied at retrieval time through a Qdrant filter, so restricted passages are never retrieved for the wrong role.
• No document text or question is written to the logs; all data stays local under data/runtime.

• Offline pytest suite split into unit, integration and security tests (access control, prompt injection and data leakage, workflow authorization).
• Two small business evaluation sets (36 and 15 cases). The repository itself notes these results are not a quality measure because the corpus is synthetic.

• Prototype on synthetic data only: not suitable for real HR data until the points listed in docs/SECURITY.md are addressed.
• The abstention gate is lexical, so distant paraphrases (for example « travail à distance » for « télétravail ») are refused: safe, but sometimes unhelpful. Making it semantic is the first roadmap item.
• No SSO, notifications, OCR for scanned PDFs or leave-balance integration; authentication is local only.

• TontynSecure web and mobile platform for managing community savings circles (tontines).
• VozelWindows dictation app that transcribes on-device and inserts text into the application you are using.