HR Chat

Last updated: 2026-05-12 Owner: Product Writer Audience: hr_admin, manager, tenant admin

Purpose

Explain how to use HR Chat as a Talent Partner to search skills, profiles and competencies, build shortlists and curate a persistent recruiter memory, always within the allowed data perimeter.

Prerequisites

• enabled role and active HR Chat capability on the tenant
• active AI service for the tenant and configured platform policy
• tenant AI panel configured by your tenant admin

Key concepts

• Talent Partner: official name of the assistant. The pill at the top of the thread shows the current model and capabilities, for example Sonnet 4.6 · memory · 8 tool.
• Recruiter memory: structured, persistent recruiter archive (saved searches, favourite candidates, active tenders). Memory is read automatically at the start of every new turn.
• Conversation memory: each session produces a synthetic summary (locked criteria, candidates shown, open questions) that stays attached to the session and is reused if you reopen it from the history.
• Agentic tools: the Talent Partner autonomously chooses up to eight tools (semantic search, candidate dossier, clarification, comparison, recommendation, validation, filter, related-skill expansion) to reach the answer. See Tool visibility below.

What you find inside the HR Chat shell

• recruiter abstract, intelligence KPIs, inline CV preview, markdown/json exports
• direct CTAs: Open candidate sheet, Open CV, move into the operational shortlist

Basic operational steps

1. Open the HR Chat module. The shell starts on the thread; you no longer choose Assistant, Search or Report: the Talent Partner decides on its own.
2. If you already have recent conversations, you will see them in the empty state. Click a row to resume the thread, or write a new request in the composer.
3. Enter a clear request, for example candidate profiles, required skills, tender context or comparison criteria.
4. While the Talent Partner works, read the execution status lines below the pill: they tell you which tool is being used.
5. The reply streams inside the same message. Below the reply you find any proposed profile cards and final operational actions.
6. For a very simple request, such as Java, expect a grounded natural reply and, only when genuinely relevant, the closest profiles inside the same message.
7. If no match exists, the chat says so explicitly and proposes adjacent searches or a broader recruiter-style cut.
8. If a profile matters, use the inline card actions: Open candidate sheet, Open CV, or move into the operational shortlist.

Quick actions on the thread

• Copy answer: hover a Talent Partner message to reveal the copy button.
• Regenerate last answer: hover the latest assistant message, press Regenerate. The system replays the same turn with the same user request.
• Edit user message: hover one of your previous messages, press Edit. An inline textarea opens with Save and regenerate or Cancel. The conversation is forked from that point.
• New chat: from the header button. The current thread, if it has messages, is auto-saved into the conversation history.

Recruiter memory

The Recruiter memory panel exposes three accordion sections that persist across sessions:

1. Saved searches — relevant queries you want to rerun or pick up later. From the empty state or the report header, use Save current search: the active session is stored as a reusable search.
2. Favourite candidates — profiles you have marked as recurring references.
3. Active tenders — persistent tender contexts the Talent Partner reuses to filter automatically.

Each section: pagination 50 + Load 50 more, click-to-confirm deletion (a first click pre-confirms, the second one deletes). Saved entries are read automatically at the start of every new turn and enrich the Talent Partner's prompt.

Conversation history

The Conversation history toggle shows:

• search bar with debounce
• list of your sessions inside the current tenant, ordered by last update
• each row: synthetic title, relative timestamp, message count, Memory active badge if the session already has a summary
• pagination Load 50 more

Clicking a session reopens it: the thread is fully reloaded, transient state (current matches, report, clarification) is reset to start clean on top of the original history.

Attaching a CV

The paperclip button in the composer lets you upload a resume as context for the current turn.

• supported formats: .pdf, .docx, .txt,
• maximum file size: 5 MB
• extracted text is capped at 200 KB in the prompt (longer CVs get truncated)
• .txt and are read locally; .pdf and .docx are parsed server-side via the /api/hr-chat/cv-parse endpoint (no LLM, plain-text extraction only)
• the final message is shaped as [Attached CV: filename]\n\n{extracted_text}\n\n{your_request}
• the Talent Partner uses the CV as context only for that turn: it is never stored as a persistent tenant asset
• not supported: scanned PDFs without selectable text (OCR), legacy .doc binaries, .docm macros

Tool visibility (tool call)

While the Talent Partner reasons, you see in real time the tool it is using, with a localised label. The eight catalog tools:

• search — semantic vector + lexical profile search
• dossier — open detailed candidate dossier
• clarify — targeted clarification request
• compare — structured comparison between candidates
• recommend — operational shortlist recommendation
• validate — validate an assumption or criterion
• filter — filter the current pool
• expand — 1-hop expansion of niche skills or certifications into adjacent catalog skills (e.g. CKA → Kubernetes, Kubernetes → Docker/Helm); results widen search recall as a soft signal, never as hard requirements (CV_ENTITY_GRAPH_01)

When a status appears it means the model is running that sub-action. You do not need to intervene: statuses update on their own and disappear as soon as the answer streams in.

Conversation language

The conversation follows the active interface language. If the workspace is in English, the Talent Partner, cards, suggestions and tool statuses are rendered in English with no manual setting.

Report workspace and shortlist workflow (downstream operations)

When you want to switch into operations, open Operational path from the compact Quick access strip. You find:

• Overview (report summary)
• Shortlist (persistent shortlist workflow with priority, notes, exclude/restore, select for comparison)
• Compare (structured comparison of selected candidates)
• Tender matrix (for tender reports: requirements/candidates matrix and per-candidate dossier)
• Archive

Export the report once the workflow is aligned. Available exports honour the canonical professional profile document engine (markdown/html/json).

Expected result

• the HR Chat shell opens directly on the thread, with a rich empty state and recent conversations when present
• the Talent Partner answers in your language, within your data perimeter, citing candidates only when truly relevant
• the tools the model is using are visible in real time
• recruiter memory is read on every turn and automatically enriches the answers
• any conversation reopened from the history keeps its context and persistent summary
• copy, regenerate, edit and new chat enable fast iteration without page reloads
• the attached CV stays local to the turn and does not pollute tenant state

Common errors

• HR Chat unavailable
• cause: role not enabled or AI service not active for the tenant plan
• solution: contact the tenant admin or platform owner

• Request too generic
• cause: a decisive detail is still missing for the ranking
• solution: answer the single clarification shown by the chat instead of rewriting the whole request

• Missing AI configuration
• cause: provider/model not configured at platform level
• solution: request intervention from a SaaS administrator

• CV not accepted
• cause: format other than .pdf / .docx / .txt / , empty file, size over 5 MB, or scanned PDF without selectable text
• solution: confirm the PDF contains selectable text (no OCR needed), or convert manually to .txt /

• CV truncated in the prompt
• cause: extracted text exceeds 200 KB
• solution: trim the CV to the relevant sections before attaching

• Recruiter memory does not seem to update
• cause: the entry was created in a different tenant or with a different user
• solution: memory is per-user per-tenant; verify Active context in the dashboard

• Conversation history empty after New chat
• cause: the previous session did not have any messages when you clicked New chat
• solution: history only lists conversations with at least one assistant exchange; resume the empty thread and ask the first question

Related guides

• roles/hr-admin
• roles/manager
• roles/admin-tenant
• ./profilo-professionale
• ./report-core

FAQ

Do I have to enable memory for each session?

No. Recruiter memory is always active and read automatically. You only decide what to save into the three slots (searches, favourites, tenders).

Is recruiter memory visible to other users in the tenant?

No. Memory is per-user inside the tenant: each recruiter keeps a private archive.

Can I send a PDF to the Talent Partner?

Yes. Since 2026-05-12 the CV attachment supports .pdf and .docx in addition to .txt and . Text extraction happens server-side. Scanned PDFs without selectable text (image-only) must be converted manually: OCR is not supported.

What happens if I close the page while the Talent Partner is replying?

The session stays saved server-side. When you come back, reopen the thread from the history to resume where you left.

Can I disable the live tool execution view?

Not selectively. Statuses appear only during the wait and vanish at the first token of the answer.